Daily Papers

During reinforcement learning with verifiable rewards (RLVR), large language models (LLMs) can exploit loopholes in their environments to obtain high rewards without improving the intended capabilities, i.e., reward hacking. Despite its risks to training efficiency and safety, monitoring and mitigating reward hacking during training remain challenging, which is limited by a lack of testbeds that reproduce hacking and reliably identify it. We introduce CATCH, a controllable testbed for studying reward hacking in coding RL. CATCH deliberately exposes environmental loopholes and provides execution-based gold labels by comparing success under a vulnerable evaluator with task correctness under an independent audit. It also can control the model's initial hacking tendency through supervised fine-tuning data mixtures and the difficulty of earning rewards through reward designing, enabling systematic comparisons of hacking dynamics and interventions. Experiments show that CATCH can produce diverse RL training trajectories with clear reward hacking, and analyses demonstrate that both initial models and reward difficulties shape the emergence of reward hacking. We further evaluate the effectiveness of different reward hacking detection and mitigation methods. A key finding is that a chain-of-thought monitor initially suppresses hacking, but this protection erodes as the policy model learn to mislead the monitor with code comments. This highlights the need to evaluate hacking mitigations throughout training with CATCH. The source code and resources are publicly released at https://github.com/THUAIS-Lab/CATCH.

Employing Large Language Models (LLMs) to assess the quality of generated responses, such as prompting instruct-tuned models or fine-tuning judge models, has become a widely adopted evaluation method. It is also known that such evaluators are vulnerable to biases, such as favoring longer responses. While it is important to overcome this problem, the specifics of these biases remain under-explored. In this work, we qualitatively identify six types of biases inherent in various judge models. We propose EvalBiasBench as a meta-evaluation collection of hand-crafted test cases for each bias type. Additionally, we present de-biasing dataset construction methods and the associated preference dataset OffsetBias. Experimental results demonstrate that fine-tuning on our dataset significantly enhances the robustness of judge models against biases and improves performance across most evaluation scenarios. We release our datasets and the fine-tuned judge model to public.

Large language models are increasingly used as chemistry assistants, yet most chemistry benchmarks still score only final answers. This masks a critical failure mode: a model may output the correct molecule, product, or option while its reasoning violates chemical logic. Existing process-level evaluators are hard to scale because LLM judges and human step-level process annotation are costly, inconsistent, and vulnerable to hallucination. We introduce ChemCoTBench-V2, a rule-verifiable diagnostic benchmark for low-cost, auditable evaluation of structured, verifier-addressable chemical reasoning traces. It spans molecular understanding, molecule editing, molecular optimization, and reaction prediction, with 5,620 evaluation samples across 18 reporting tasks. Models must expose key intermediate steps in expert-designed templates, and those steps are checked with deterministic chemistry rules and, for closed-answer tasks, reference traces rather than another LLM judge. Open-ended molecular optimization is evaluated with oracle-verifiable state constraints rather than strict trace matching. The benchmark reports three separate signals: final-answer correctness, template adherence, and step-wise verifier correctness over expert-refined intermediate commitments. Experiments on frontier models reveal a persistent gap between final-answer success and structured-reasoning-state consistency: models often follow the requested format while failing chemical-step checks, or answer correctly with weak supporting reasoning. ChemCoTBench-V2 enables fine-grained model comparison and identifies the concrete step at which the trace first violates the verifier.

The remarkable capabilities of large language models (LLMs) in language understanding and generation have not rendered them immune to hallucinations. LLMs can still generate plausible-sounding but factually incorrect or fabricated information. As LLM-empowered chatbots become popular, laypeople may frequently ask health-related queries and risk falling victim to these LLM hallucinations, resulting in various societal and healthcare implications. In this work, we conduct a pioneering study of hallucinations in LLM-generated responses to real-world healthcare queries from patients. We propose MedHalu, a carefully crafted first-of-its-kind medical hallucination dataset with a diverse range of health-related topics and the corresponding hallucinated responses from LLMs with labeled hallucination types and hallucinated text spans. We also introduce MedHaluDetect framework to evaluate capabilities of various LLMs in detecting hallucinations. We also employ three groups of evaluators -- medical experts, LLMs, and laypeople -- to study who are more vulnerable to these medical hallucinations. We find that LLMs are much worse than the experts. They also perform no better than laypeople and even worse in few cases in detecting hallucinations. To fill this gap, we propose expert-in-the-loop approach to improve hallucination detection through LLMs by infusing expert reasoning. We observe significant performance gains for all the LLMs with an average macro-F1 improvement of 6.3 percentage points for GPT-4.

Almost every paper on LLM jailbreaks and prompt injection reports an attack-success rate (ASR), and that number is assigned not by people but by an automated judge: either a safety classifier trained for the task, or a general chat model prompted to grade. The judge is rarely checked. We check it. Using 596 human-labeled completions from the HarmBench classifier validation set, we compare the two judge families against human majority votes and then attack them. The two families fail in opposite ways. The dedicated classifier over-flags (precision 0.835, recall 0.974); three different LLM-as-judges keep high precision (0.81 to 0.94) but show erratic recall (0.06 to 0.65), so the same responses produce very different ASR depending on which judge scores them. The two families also differ sharply in robustness. Wrappers that leave the harmful text untouched and only add benign framing flip every LLM-judge between 57% and 100% of the time, and a single prepended refusal sentence accounts for much of this (39% to 88%). The dedicated classifier resists these surface attacks (at most 6.7%), but a white-box GCG attack on its open weights flips 70% of confident true positives (21 of 30; 95% CI 54 to 86%) even at a small optimization budget. A two-annotator audit confirms the attacks leave the harm intact: every one of 80 sampled flips still contained the harmful content. Because a large and growing share of reported ASR comes from LLM-judges, many such numbers are unreliable both on average and under deliberate pressure. We recommend that papers report judge precision and recall on a human-labeled slice, report ASR corrected for judge precision, and include an adversarial check of the judge. Our code is released.

Large Language Models (LLMs) are increasingly relied upon to evaluate text outputs of other LLMs, thereby influencing leaderboards and development decisions. However, concerns persist over the accuracy of these assessments and the potential for misleading conclusions. In this work, we investigate the effectiveness of LLMs as evaluators for text generation tasks. We propose FBI, a novel framework designed to examine the proficiency of Evaluator LLMs in assessing four critical abilities in other LLMs: factual accuracy, instruction following, coherence in long-form writing, and reasoning proficiency. By introducing targeted perturbations in answers generated by LLMs, that clearly impact one of these key capabilities, we test whether an Evaluator LLM can detect these quality drops. By creating a total of 2400 perturbed answers covering 22 perturbation categories, we conduct a comprehensive study using different evaluation strategies on five prominent LLMs commonly used as evaluators in the literature. Our findings reveal significant shortcomings in current Evaluator LLMs, which failed to identify quality drops in over 50\% of cases on average. Single-answer and pairwise evaluations demonstrated notable limitations, whereas reference-based evaluations showed comparatively better performance. These results underscore the unreliable nature of current Evaluator LLMs and advocate for cautious implementation in practical applications. Code and data are available at https://github.com/AI4Bharat/FBI.

Large language models (LLMs) are increasingly deployed in automated code-review systems, where their approvals can determine which code is merged into shared repositories. However, it is unclear whether review agents can detect vulnerability-introducing code when an attacker controls both the code change and the persuasive Pull Request (PR) narrative designed to mask it. We introduce SEVRA-BENCH (Social Engineering of Vulnerabilities in Review Agents), a benchmark that measures how often a review agent approves such adversarial PR s. Each PR in SEVRA-BENCH is built from a historical commit that fixed a vulnerability. We automatically reverse that fix to extract the original vulnerable code, and submit the resulting code change as a PR wrapped in one of 15 social-engineering framings. To test review-agent resilience to narrative manipulation, these framings vary dimensions such as supporting evidence, conveyed urgency, signals of prior approval, and appeals to authority. SEVRA-BENCH evaluates a retained challenge split of roughly 1000 adversarial PRs drawn from publicly disclosed vulnerability fixes across the top 10 entries of the MITRE's 2025 most dangerous software weaknesses. Evaluating 8 review agents against this benchmark, we reveal that review agents are susceptible to narrative manipulation, exposing a significant gap in security capabilities.

Benchmark contamination has become a significant concern in the LLM evaluation community. Previous Agents-as-an-Evaluator address this issue by involving agents in the generation of questions. Despite their success, the biases in Agents-as-an-Evaluator methods remain largely unexplored. In this paper, we present a theoretical formulation of evaluation bias, providing valuable insights into designing unbiased evaluation protocols. Furthermore, we identify two type of bias in Agents-as-an-Evaluator through carefully designed probing tasks on a minimal Agents-as-an-Evaluator setup. To address these issues, we propose the Unbiased Evaluator, an evaluation protocol that delivers a more comprehensive, unbiased, and interpretable assessment of LLMs.Extensive experiments reveal significant room for improvement in current LLMs. Additionally, we demonstrate that the Unbiased Evaluator not only offers strong evidence of benchmark contamination but also provides interpretable evaluation results.

Trustworthy capability evaluations are crucial for ensuring the safety of AI systems, and are becoming a key component of AI regulation. However, the developers of an AI system, or the AI system itself, may have incentives for evaluations to understate the AI's actual capability. These conflicting interests lead to the problem of sandbagging, which we define as strategic underperformance on an evaluation. In this paper we assess sandbagging capabilities in contemporary language models (LMs). We prompt frontier LMs, like GPT-4 and Claude 3 Opus, to selectively underperform on dangerous capability evaluations, while maintaining performance on general (harmless) capability evaluations. Moreover, we find that models can be fine-tuned, on a synthetic dataset, to hide specific capabilities unless given a password. This behaviour generalizes to high-quality, held-out benchmarks such as WMDP. In addition, we show that both frontier and smaller models can be prompted or password-locked to target specific scores on a capability evaluation. We have mediocre success in password-locking a model to mimic the answers a weaker model would give. Overall, our results suggest that capability evaluations are vulnerable to sandbagging. This vulnerability decreases the trustworthiness of evaluations, and thereby undermines important safety decisions regarding the development and deployment of advanced AI systems.

The rapid integration of Large Language Models (LLMs) into high-stakes domains necessitates reliable safety and compliance evaluation. However, existing static benchmarks are ill-equipped to address the dynamic nature of AI risks and evolving regulations, creating a critical safety gap. This paper introduces a new paradigm of agentic safety evaluation, reframing evaluation as a continuous and self-evolving process rather than a one-time audit. We then propose a novel multi-agent framework SafeEvalAgent, which autonomously ingests unstructured policy documents to generate and perpetually evolve a comprehensive safety benchmark. SafeEvalAgent leverages a synergistic pipeline of specialized agents and incorporates a Self-evolving Evaluation loop, where the system learns from evaluation results to craft progressively more sophisticated and targeted test cases. Our experiments demonstrate the effectiveness of SafeEvalAgent, showing a consistent decline in model safety as the evaluation hardens. For instance, GPT-5's safety rate on the EU AI Act drops from 72.50% to 36.36% over successive iterations. These findings reveal the limitations of static assessments and highlight our framework's ability to uncover deep vulnerabilities missed by traditional methods, underscoring the urgent need for dynamic evaluation ecosystems to ensure the safe and responsible deployment of advanced AI.

Assessing cybersecurity vulnerability awareness in coding agents requires evaluations that reveal capability gaps and remain informative as models evolve. Static benchmarks offer fixed coverage and difficulty, while scarce vulnerable repositories and costly expert authoring limit their renewal at scale. We introduce SecProbe, a framework for adaptive evaluation that combines Item Response Theory (IRT) with on-demand synthesis of repository-scale vulnerability-repair tasks. From observed performance, SecProbe estimates agent ability and identifies where additional evidence is most informative, selecting existing tasks or synthesizing new ones accordingly. As one use case, we construct 353 tasks spanning six programming languages and 151 CWE types and evaluate nine frontier models with two agent harnesses. Success rates peak at 28.33\%, highlighting substantial gaps in vulnerability recognition and repair. Compared with random and one-shot baselines, SecProbe achieves comparable agent ability estimates while requiring agents to solve up to 29.5\% fewer tasks. These results support adaptive evaluation as an efficient and discriminative approach to assessing cybersecurity vulnerability awareness.

Automated LLM-as-a-Judge frameworks have become the de facto standard for scalable evaluation across natural language processing. For instance, in safety evaluation, these judges are relied upon to evaluate harmfulness in order to benchmark the robustness of safety against adversarial attacks. However, we show that existing validation protocols fail to account for substantial distribution shifts inherent to red-teaming: diverse victim models exhibit distinct generation styles, attacks distort output patterns, and semantic ambiguity varies significantly across jailbreak scenarios. Through a comprehensive audit using 6642 human-verified labels, we reveal that the unpredictable interaction of these shifts often causes judge performance to degrade to near random chance. This stands in stark contrast to the high human agreement reported in prior work. Crucially, we find that many attacks inflate their success rates by exploiting judge insufficiencies rather than eliciting genuinely harmful content. To enable more reliable evaluation, we propose ReliableBench, a benchmark of behaviors that remain more consistently judgeable, and JudgeStressTest, a dataset designed to expose judge failures. Data available at: https://github.com/SchwinnL/LLMJudgeReliability.

Despite the rapid advancement of LLM-based agents, the reliable evaluation of their safety and security remains a significant challenge. Existing rule-based or LLM-based evaluators often miss dangers in agents' step-by-step actions, overlook subtle meanings, fail to see how small issues compound, and get confused by unclear safety or security rules. To overcome this evaluation crisis, we introduce AgentAuditor, a universal, training-free, memory-augmented reasoning framework that empowers LLM evaluators to emulate human expert evaluators. AgentAuditor constructs an experiential memory by having an LLM adaptively extract structured semantic features (e.g., scenario, risk, behavior) and generate associated chain-of-thought reasoning traces for past interactions. A multi-stage, context-aware retrieval-augmented generation process then dynamically retrieves the most relevant reasoning experiences to guide the LLM evaluator's assessment of new cases. Moreover, we developed ASSEBench, the first benchmark designed to check how well LLM-based evaluators can spot both safety risks and security threats. ASSEBench comprises 2293 meticulously annotated interaction records, covering 15 risk types across 29 application scenarios. A key feature of ASSEBench is its nuanced approach to ambiguous risk situations, employing "Strict" and "Lenient" judgment standards. Experiments demonstrate that AgentAuditor not only consistently improves the evaluation performance of LLMs across all benchmarks but also sets a new state-of-the-art in LLM-as-a-judge for agent safety and security, achieving human-level accuracy. Our work is openly accessible at https://github.com/Astarojth/AgentAuditor.

AI systems can fail silently. The failure propagates through training loops, evaluation pipelines, and production monitoring stacks until downstream harm makes it visible. This paper introduces evaluation blindness: a measurement function M exhibits evaluation blindness with respect to failure class F when it produces readings indistinguishable from a healthy state while the system is actually failing, with no auxiliary signal flagging the gap. The problem surfaces at two lifecycle stages the literature has treated separately. At training time, reward models are gamed, importance-sampling corrections are silently miscalculated, and benchmark contamination inflates fine-tuning evaluations, all while loss curves look healthy and gradient updates proceed normally. At deployment time, monitoring fails to catch six classes of production failure, including an Operational category that is 100% silent by structural definition. We provide a formal detectability predicate unifying both stages. Four training-time case studies trace concrete breakdowns, including a real implementation bug in TRL PR #6594 where gradients are corrupted as loss decreases normally. A six-class taxonomy validated against 50 real-world incidents from court documents and regulatory filings finds that 53% of verifiable public failures were silent. A failure budget framework ties acceptable failure rates to use-case risk class. The implication is direct: measurement infrastructure is a correctness concern across the full AI lifecycle, not just at evaluation time. Data, code, and taxonomy schema are at https://github.com/priyanka25aug/llm-failure-taxonomy.

Language models can distinguish between testing and deployment phases -- a capability known as evaluation awareness. This has significant safety and policy implications, potentially undermining the reliability of evaluations that are central to AI governance frameworks and voluntary industry commitments. In this paper, we study evaluation awareness in Llama-3.3-70B-Instruct. We show that linear probes can separate real-world evaluation and deployment prompts, suggesting that current models internally represent this distinction. We also find that current safety evaluations are correctly classified by the probes, suggesting that they already appear artificial or inauthentic to models. Our findings underscore the importance of ensuring trustworthy evaluations and understanding deceptive capabilities. More broadly, our work showcases how model internals may be leveraged to support blackbox methods in safety audits, especially for future models more competent at evaluation awareness and deception.

Existing LLM safety evaluations rely on binary attack-success rates and domain-agnostic taxonomies, leaving regulated Banking, Financial Services, and Insurance (BFSI) deployments exposed to failures elicited through legally or professionally plausible framing. We introduce RAHS (Risk-Adjusted Harm Score), a risk-sensitive metric jointly capturing disclosure severity, disclaimer mitigation, and inter-judge agreement, and FinRedTeamBench, a 989-prompt benchmark spanning seven BFSI risk areas and 34 sub-categories mapped to regulatory frameworks. Evaluation uses an ensemble of three heterogeneous LLM judges, validated against human experts, and an adaptive multi-turn red-teaming pipeline. On nine open-weight models, RAHS preserves separation under near-ceiling ASR, ranking is stable under hyperparameter sweeps, and multi-turn pressure drives not only more jailbreaks but more operationally severe disclosures, exposing failure modes that single-turn, domain-agnostic evaluations cannot reveal.

Text-to-image (T2I) models have rapidly advanced, enabling the generation of high-quality images from text prompts across various domains. However, these models present notable safety concerns, including the risk of generating harmful, biased, or private content. Current research on assessing T2I safety remains in its early stages. While some efforts have been made to evaluate models on specific safety dimensions, many critical risks remain unexplored. To address this gap, we introduce T2ISafety, a safety benchmark that evaluates T2I models across three key domains: toxicity, fairness, and bias. We build a detailed hierarchy of 12 tasks and 44 categories based on these three domains, and meticulously collect 70K corresponding prompts. Based on this taxonomy and prompt set, we build a large-scale T2I dataset with 68K manually annotated images and train an evaluator capable of detecting critical risks that previous work has failed to identify, including risks that even ultra-large proprietary models like GPTs cannot correctly detect. We evaluate 12 prominent diffusion models on T2ISafety and reveal several concerns including persistent issues with racial fairness, a tendency to generate toxic content, and significant variation in privacy protection across the models, even with defense methods like concept erasing. Data and evaluator are released under https://github.com/adwardlee/t2i_safety.

Large Language Models (LLMs) have demonstrated exceptional capabilities across diverse tasks, driving the development and widespread adoption of LLM-as-a-Judge systems for automated evaluation, including red teaming and benchmarking. However, these systems are susceptible to adversarial attacks that can manipulate evaluation outcomes, raising critical concerns about their robustness and trustworthiness. Existing evaluation methods for LLM-based judges are often fragmented and lack a unified framework for comprehensive robustness assessment. Furthermore, the impact of prompt template design and model selection on judge robustness has rarely been explored, and their performance in real-world deployments remains largely unverified. To address these gaps, we introduce RobustJudge, a fully automated and scalable framework designed to systematically evaluate the robustness of LLM-as-a-Judge systems. Specifically, RobustJudge investigates the effectiveness of 15 attack methods and 7 defense strategies across 12 models (RQ1), examines the impact of prompt template design and model selection (RQ2), and evaluates the security of real-world deployments (RQ3). Our study yields three key findings: (1) LLM-as-a-Judge systems are highly vulnerable to attacks such as PAIR and combined attacks, while defense mechanisms such as re-tokenization and LLM-based detectors can provide enhanced protection; (2) robustness varies substantially across prompt templates (up to 40%); (3) deploying RobustJudge on Alibaba's PAI platform uncovers previously undiscovered vulnerabilities. These results offer practical insights for building trustworthy LLM-as-a-Judge systems.

Large Language Models have gained considerable attention for their revolutionary capabilities. However, there is also growing concern on their safety implications, making a comprehensive safety evaluation for LLMs urgently needed before model deployment. In this work, we propose S-Eval, a new comprehensive, multi-dimensional and open-ended safety evaluation benchmark. At the core of S-Eval is a novel LLM-based automatic test prompt generation and selection framework, which trains an expert testing LLM Mt combined with a range of test selection strategies to automatically construct a high-quality test suite for the safety evaluation. The key to the automation of this process is a novel expert safety-critique LLM Mc able to quantify the riskiness score of a LLM's response, and additionally produce risk tags and explanations. Besides, the generation process is also guided by a carefully designed risk taxonomy with four different levels, covering comprehensive and multi-dimensional safety risks of concern. Based on these, we systematically construct a new and large-scale safety evaluation benchmark for LLMs consisting of 220,000 evaluation prompts, including 20,000 base risk prompts (10,000 in Chinese and 10,000 in English) and 200, 000 corresponding attack prompts derived from 10 popular adversarial instruction attacks against LLMs. Moreover, considering the rapid evolution of LLMs and accompanied safety threats, S-Eval can be flexibly configured and adapted to include new risks, attacks and models. S-Eval is extensively evaluated on 20 popular and representative LLMs. The results confirm that S-Eval can better reflect and inform the safety risks of LLMs compared to existing benchmarks. We also explore the impacts of parameter scales, language environments, and decoding parameters on the evaluation, providing a systematic methodology for evaluating the safety of LLMs.

The advent of instruction-tuned Large Language Models designed for coding tasks (Code LLMs) has transformed software engineering practices. However, their robustness against various input challenges remains a critical concern. This study introduces DegradePrompter, a novel method designed to systematically evaluate the robustness of instruction-tuned Code LLMs. We assess the impact of diverse input challenges on the functionality and correctness of generated code using rigorous metrics and established benchmarks. Our comprehensive evaluation includes five state-of-the-art open-source models and three production-grade closed-source models, revealing varying degrees of robustness. Open-source models demonstrate an increased susceptibility to input perturbations, resulting in declines in functional correctness ranging from 12% to 34%. In contrast, commercial models demonstrate relatively greater resilience, with performance degradation ranging from 3% to 24%. To enhance the robustness of the models against these vulnerabilities, we investigate a straightforward yet effective mitigation strategy. Our findings highlight the need for robust defense mechanisms and comprehensive evaluations during both the development and deployment phases to ensure the resilience and reliability of automated code generation systems.

As large language models (LLMs) increasingly integrate native code interpreters, they enable powerful real-time execution capabilities, substantially expanding their utility. However, such integrations introduce potential system-level cybersecurity threats, fundamentally different from prompt-based vulnerabilities. To systematically evaluate these interpreter-specific risks, we propose CIRCLE (Code-Interpreter Resilience Check for LLM Exploits), a simple benchmark comprising 1,260 prompts targeting CPU, memory, and disk resource exhaustion. Each risk category includes explicitly malicious ("direct") and plausibly benign ("indirect") prompt variants. Our automated evaluation framework assesses not only whether LLMs refuse or generates risky code, but also executes the generated code within the interpreter environment to evaluate code correctness, simplifications made by the LLM to make the code safe, or execution timeouts. Evaluating 7 commercially available models from OpenAI and Google, we uncover significant and inconsistent vulnerabilities. For instance, evaluations show substantial disparities even within providers - OpenAI's o4-mini correctly refuses risky requests at 7.1%, notably higher rates compared to GPT-4.1 at 0.5%. Results particularly underscore that indirect, socially-engineered prompts substantially weaken model defenses. This highlights an urgent need for interpreter-specific cybersecurity benchmarks, dedicated mitigation tools (e.g., guardrails), and clear industry standards to guide safe and responsible deployment of LLM interpreter integrations. The benchmark dataset and evaluation code are publicly released to foster further research.

LLM agents that invoke privileged tools are vulnerable to indirect prompt injection (IPI), in which adversarial instructions embedded in retrieved data hijack the agent's actions. A growing body of work evaluates defenses against IPI, but the validity of that evaluation is rarely examined. We audit an IPI benchmark and its harness and identify four defect classes -- silent payload non-delivery, attack success scored by tool identity rather than arguments, false-rejection rate conflated with model incapacity, and the absence of an audit trail -- each of which yields a plausible, publishable, and incorrect number. We quantify the distortion by re-scoring identical execution traces under the defective and corrected definitions: on real agent behaviour, the tool-identity scorer reports a 21.7% attack-success rate where the true argument-level rate is 1.2%. In the sharpest case, an open model previously reported at 62.8% registers 0% under the corrected harness -- the prior figure largely an artifact of undelivered payloads and identity-level scoring. We release a harness whose construction makes each defect unrepresentable -- machine-checkable payload placement, argument-level attacker predicates, per-scenario environments, and mandatory trace persistence -- and use it to report three quantities the field does not: whether a compromised agent discloses the attack, the full security/utility operating curve of an LLM-judge defense, and tool-calling capability disentangled from defensive over-blocking. A corrected harness further overturns a reported "capability barrier": a model deemed incapable of tool use is in fact fully capable, its earlier result an artifact of environment mismatch. We argue that evaluation validity is a prerequisite for, not a footnote to, defense claims in agentic security, and provide an instrument that enforces it.

Large Language Models (LLMs) are increasingly employed as automated evaluators to assess the safety of generated content, yet their reliability in this role remains uncertain. This study evaluates a diverse set of 11 LLM judge models across critical safety domains, examining three key aspects: self-consistency in repeated judging tasks, alignment with human judgments, and susceptibility to input artifacts such as apologetic or verbose phrasing. Our findings reveal that biases in LLM judges can significantly distort the final verdict on which content source is safer, undermining the validity of comparative evaluations. Notably, apologetic language artifacts alone can skew evaluator preferences by up to 98\%. Contrary to expectations, larger models do not consistently exhibit greater robustness, while smaller models sometimes show higher resistance to specific artifacts. To mitigate LLM evaluator robustness issues, we investigate jury-based evaluations aggregating decisions from multiple models. Although this approach both improves robustness and enhances alignment to human judgements, artifact sensitivity persists even with the best jury configurations. These results highlight the urgent need for diversified, artifact-resistant methodologies to ensure reliable safety assessments.

Proprietary LMs such as GPT-4 are often employed to assess the quality of responses from various LMs. However, concerns including transparency, controllability, and affordability strongly motivate the development of open-source LMs specialized in evaluations. On the other hand, existing open evaluator LMs exhibit critical shortcomings: 1) they issue scores that significantly diverge from those assigned by humans, and 2) they lack the flexibility to perform both direct assessment and pairwise ranking, the two most prevalent forms of assessment. Additionally, they do not possess the ability to evaluate based on custom evaluation criteria, focusing instead on general attributes like helpfulness and harmlessness. To address these issues, we introduce Prometheus 2, a more powerful evaluator LM than its predecessor that closely mirrors human and GPT-4 judgements. Moreover, it is capable of processing both direct assessment and pair-wise ranking formats grouped with a user-defined evaluation criteria. On four direct assessment benchmarks and four pairwise ranking benchmarks, Prometheus 2 scores the highest correlation and agreement with humans and proprietary LM judges among all tested open evaluator LMs. Our models, code, and data are all publicly available at https://github.com/prometheus-eval/prometheus-eval.

AI agents are entering high-risk production settings, where they use tools, retain context, follow policies, handle private data, and interact with users over multiple turns. Yet many evaluation methods still judge isolated outputs or static tasks, missing failures that emerge through trajectory, pressure, and adversarial interaction. We introduce ProofAgent Harness, open infrastructure for scalable, auditable, and adversarial AI agent evaluation. The harness provides evaluation infrastructure around an agent: it curates evaluation intelligence, runs adversarial multi-turn trials, captures behavioral traces, applies post-hoc multi-juror scoring, resolves disagreement, and produces evidence-linked reports. Its open design allows developers and researchers to extend domains, traps, metrics, juror personas, scoring rules, and reporting formats. At its core is Adversarial Multi-Juror Scoring with Turn-Level Audit, which evaluates completed agent behavior under pressure using calibrated juror personas, consensus checks, and turn-level evidence. Experiments across customer support, medical triage, privacy and security, and code generation agents show that strong agents fail selectively through weak metrics, fragile turns, unsafe reframing, and manipulation paths. We also find that a small quantized local Harness LLM can challenge production agents powered by best-in-class large LLMs, suggesting that evaluation capability emerges from the full harness pipeline rather than model scale alone. ProofAgent Harness turns AI agent evaluation from a static score into scalable adversarial evaluation infrastructure: repeatable, evidence-backed, extensible, and actionable before deployment.

An ideal model evaluation should achieve two goals: identifying where the model fails and providing actionable improvement guidance. Toward these goals for Language Model (LM) evaluations, we formulate the problem of generating a weakness profile, a set of weaknesses expressed in natural language, given an LM's performance on every individual instance in a benchmark. We introduce a suite of quantitative assessments to compare different weakness profiling methods. We also propose a weakness profiling method EvalTree. It constructs a capability tree where each node represents a capability described in natural language and is linked to a subset of benchmark instances that specifically evaluate this capability; it then extracts nodes where the LM performs poorly to generate a weakness profile. On the MATH and WildChat benchmarks, we show that EvalTree outperforms baseline weakness profiling methods by identifying weaknesses more precisely and comprehensively. Weakness profiling further enables weakness-guided data collection, and training data collection guided by EvalTree-identified weaknesses improves LM performance more than other data collection strategies. We also show how EvalTree exposes flaws in Chatbot Arena's human-voter-based evaluation practice. To facilitate future work, we release our code and an interface that allows practitioners to interactively explore the capability trees built by EvalTree.

Evaluating aligned large language models' (LLMs) ability to recognize and reject unsafe user requests is crucial for safe, policy-compliant deployments. Existing evaluation efforts, however, face three limitations that we address with SORRY-Bench, our proposed benchmark. First, existing methods often use coarse-grained taxonomies of unsafe topics, and are over-representing some fine-grained topics. For example, among the ten existing datasets that we evaluated, tests for refusals of self-harm instructions are over 3x less represented than tests for fraudulent activities. SORRY-Bench improves on this by using a fine-grained taxonomy of 45 potentially unsafe topics, and 450 class-balanced unsafe instructions, compiled through human-in-the-loop methods. Second, linguistic characteristics and formatting of prompts are often overlooked, like different languages, dialects, and more -- which are only implicitly considered in many evaluations. We supplement SORRY-Bench with 20 diverse linguistic augmentations to systematically examine these effects. Third, existing evaluations rely on large LLMs (e.g., GPT-4) for evaluation, which can be computationally expensive. We investigate design choices for creating a fast, accurate automated safety evaluator. By collecting 7K+ human annotations and conducting a meta-evaluation of diverse LLM-as-a-judge designs, we show that fine-tuned 7B LLMs can achieve accuracy comparable to GPT-4 scale LLMs, with lower computational cost. Putting these together, we evaluate over 40 proprietary and open-source LLMs on SORRY-Bench, analyzing their distinctive refusal behaviors. We hope our effort provides a building block for systematic evaluations of LLMs' safety refusal capabilities, in a balanced, granular, and efficient manner.

Automated vulnerability detection research has made substantial progress, yet its real-world impact remains limited. Prior work found that current vulnerability datasets suffer from issues including label inaccuracy rates of 20%-71%, extensive duplication, and poor coverage of critical Common Weakness Enumeration (CWE). These issues create a significant generalization gap where models achieve misleading In-Distribution (ID) accuracies (testing on splits from the same dataset) by exploiting spurious correlations rather than learning true vulnerability patterns. To address these limitations, we present a three-part solution. First, we introduce BenchVul, which is a manually curated and balanced test dataset covering the MITRE Top 25 Most Dangerous CWEs, to enable fair model evaluation. Second, we construct a high-quality training dataset, TitanVul, comprising 38,548 functions by aggregating seven public sources and applying deduplication and validation using a novel multi-agent LLM pipeline. Third, we propose a Realistic Vulnerability Generation (RVG) pipeline, which synthesizes context-aware vulnerability examples for underrepresented but critical CWE types through simulated development workflows. Our evaluation reveals that In-Distribution (ID) performance does not reliably predict Out-of-Distribution (OOD) performance on BenchVul. For example, a model trained on BigVul achieves the highest 0.703 ID accuracy but fails on BenchVul's real-world samples (0.493 OOD accuracy). Conversely, a model trained on our TitanVul achieves the highest OOD performance on both the real-world (0.881) and synthesized (0.785) portions of BenchVul, improving upon the next-best performing dataset by 5.3% and 11.8% respectively, despite a modest ID score (0.590). Augmenting TitanVul with our RVG further boosts this leading OOD performance, improving accuracy on real-world data by 5.8% (to 0.932).

Large Language Models (LLMs) have significantly aided developers by generating or assisting in code writing, enhancing productivity across various tasks. While identifying incorrect code is often straightforward, detecting vulnerabilities in functionally correct code is more challenging, especially for developers with limited security knowledge, which poses considerable security risks of using LLM-generated code and underscores the need for robust evaluation benchmarks that assess both functional correctness and security. Current benchmarks like CyberSecEval and SecurityEval attempt to solve it but are hindered by unclear and impractical specifications, failing to assess both functionality and security accurately. To tackle these deficiencies, we introduce CWEval, a novel outcome-driven evaluation framework designed to enhance the evaluation of secure code generation by LLMs. This framework not only assesses code functionality but also its security simultaneously with high-quality task specifications and outcome-driven test oracles which provides high accuracy. Coupled with CWEval-bench, a multilingual, security-critical coding benchmark, CWEval provides a rigorous empirical security evaluation on LLM-generated code, overcoming previous benchmarks' shortcomings. Through our evaluations, CWEval reveals a notable portion of functional but insecure code produced by LLMs, and shows a serious inaccuracy of previous evaluations, ultimately contributing significantly to the field of secure code generation. We open-source our artifact at: https://github.com/Co1lin/CWEval .

Current evaluations of LLM safety predominantly rely on severity-based taxonomies to assess the harmfulness of malicious queries. We argue that this formulation requires re-examination as it assumes uniform risk across all malicious queries, neglecting Execution Likelihood--the conditional probability of a threat being realized given the model's response. In this work, we introduce Expected Harm, a metric that weights the severity of a jailbreak by its execution likelihood, modeled as a function of execution cost. Through empirical analysis of state-of-the-art models, we reveal a systematic Inverse Risk Calibration: models disproportionately exhibit stronger refusal behaviors for low-likelihood (high-cost) threats while remaining vulnerable to high-likelihood (low-cost) queries. We demonstrate that this miscalibration creates a structural vulnerability: by exploiting this property, we increase the attack success rate of existing jailbreaks by up to 2times. Finally, we trace the root cause of this failure using linear probing, which reveals that while models encode severity in their latent space to drive refusal decisions, they possess no distinguishable internal representation of execution cost, making them "blind" to this critical dimension of risk.

Large language models (LLMs) can often tell when they are being tested. This ability, known as evaluation awareness, can lead models to act more cautiously during safety evaluations than in real use. Existing work has studied evaluation awareness almost entirely in English, although models serve users in many languages. Indic languages are spoken by over a billion people, yet evaluation awareness remains almost entirely unstudied in them. We test four instruction-tuned models from three families in five Indic languages alongside English. A set of cues varies whether the model is told that it is monitored and whether the monitoring is a safety test. Models refuse harmful requests more often when they are told that they are being tested. This increase appears in all four models and all six languages. In Indic languages, refusal of harmful requests rises by 14.4 percentage points on average. The test cue also brings Indic refusal of harmful requests level with English, but raises refusal of benign Indic requests from 6.9% to 11.3%. We release all 47,880 judged responses to support further work. These results suggest that safety evaluations in Indic languages may not reflect how models treat their everyday users.

Powered by remarkable advancements in Large Language Models (LLMs), Multimodal Large Language Models (MLLMs) demonstrate impressive capabilities in manifold tasks. However, the practical application scenarios of MLLMs are intricate, exposing them to potential malicious instructions and thereby posing safety risks. While current benchmarks do incorporate certain safety considerations, they often lack comprehensive coverage and fail to exhibit the necessary rigor and robustness. For instance, the common practice of employing GPT-4V as both the evaluator and a model to be evaluated lacks credibility, as it tends to exhibit a bias toward its own responses. In this paper, we present MLLMGuard, a multidimensional safety evaluation suite for MLLMs, including a bilingual image-text evaluation dataset, inference utilities, and a lightweight evaluator. MLLMGuard's assessment comprehensively covers two languages (English and Chinese) and five important safety dimensions (Privacy, Bias, Toxicity, Truthfulness, and Legality), each with corresponding rich subtasks. Focusing on these dimensions, our evaluation dataset is primarily sourced from platforms such as social media, and it integrates text-based and image-based red teaming techniques with meticulous annotation by human experts. This can prevent inaccurate evaluation caused by data leakage when using open-source datasets and ensures the quality and challenging nature of our benchmark. Additionally, a fully automated lightweight evaluator termed GuardRank is developed, which achieves significantly higher evaluation accuracy than GPT-4. Our evaluation results across 13 advanced models indicate that MLLMs still have a substantial journey ahead before they can be considered safe and responsible.

Backdoor poisoning attacks add poisoned examples to otherwise-clean finetuning data, pairing a trigger with a target behavior that the model learns to produce when the trigger appears. Existing evaluations typically fix the number of poisoned examples and sample them at random from a candidate pool. We show that this can severely underestimate worst-case vulnerability: across three LLaMA-3-8B backdoor settings, holding the model, clean data, and poison count fixed, attack success ranges from 3% to 80% depending only on which poison set is chosen. We formalize poison selection as oracle-budgeted set optimization and introduce SAILS (Set-level Audit-Informed Iterative Learned Selection), which learns a set scorer from a few hundred finetune-and-evaluate runs, ranks millions of candidate sets, and audits only a small shortlist. SAILS improves held-out attack success by 30 percentage points on average over the strongest influence baselines, transfers from small-scale to full-scale finetuning, and extends to code-generation, agentic, and API-only backdoors.

Large language models (LLMs) are increasingly deployed in medical contexts, raising critical concerns about safety, alignment, and susceptibility to adversarial manipulation. While prior benchmarks assess model refusal capabilities for harmful prompts, they often lack clinical specificity, graded harmfulness levels, and coverage of jailbreak-style attacks. We introduce CARES (Clinical Adversarial Robustness and Evaluation of Safety), a benchmark for evaluating LLM safety in healthcare. CARES includes over 18,000 prompts spanning eight medical safety principles, four harm levels, and four prompting styles: direct, indirect, obfuscated, and role-play, to simulate both malicious and benign use cases. We propose a three-way response evaluation protocol (Accept, Caution, Refuse) and a fine-grained Safety Score metric to assess model behavior. Our analysis reveals that many state-of-the-art LLMs remain vulnerable to jailbreaks that subtly rephrase harmful prompts, while also over-refusing safe but atypically phrased queries. Finally, we propose a mitigation strategy using a lightweight classifier to detect jailbreak attempts and steer models toward safer behavior via reminder-based conditioning. CARES provides a rigorous framework for testing and improving medical LLM safety under adversarial and ambiguous conditions.

Evaluations should do more than measure a models current performance. They should tell us what to fix for the next model iteration and provide a way to generate targeted post training data. Most evaluation pipelines identify weak examples, topics, or categories, but they leave the underlying capability failure implicit: they say where a model fails, not why. We introduce CRAFT, a method that converts any rubric based evaluation dataset into a model specific diagnosis of weak capabilities. CRAFT treats each grading criterion as a capability probe: it extracts a capability description from every prompt rubric pair, clusters these descriptions into a hierarchical capability tree, scores the target model at every node, and selects low performing nodes dynamically across tree levels, at the granularity where each failure is clearest. The selected weak capabilities then direct the generation of targeted supervised finetuning data. Holding the data generation, finetuning, and evaluation setup fixed, we compare CRAFT against prompt level EvalTree clustering and untargeted random generation on four open source models, two professional domains (finance and legal), and 13 held out benchmarks disjoint from the diagnostic data. CRAFT achieves the strongest finance domain average for all four models under repeated temperature decoding; on legal domain, it is strongest for three of four models and remains within the decoding variance bands of the best baseline on the fourth. Diagnosing weaknesses at the level of rubric criteria, rather than prompts or categories, thus yields both a sharper picture of what a model cannot do and measurably better models after finetuning on that diagnosis.

Agent evaluation requires assessing complex multi-step behaviors involving tool use and intermediate reasoning, making it costly and expertise-intensive. A natural question arises: can frontier coding assistants reliably automate this evaluation process? Our study shows that simply prompting coding assistants is insufficient for this task. Without domain-specific evaluation knowledge, frontier coding assistants achieve only a 30% execution success rate and produce over-engineered evaluations averaging 12+ metrics per agent, indicating that strong coding ability does not automatically translate to reliable agent evaluation. We introduce EvalAgent, an AI assistant that automates the end-to-end agent evaluation pipeline. EvalAgent encodes evaluation domain expertise as evaluation skills (procedural instructions, reusable code and templates, and dynamically retrieved API documentation) that compose into a trace-based pipeline producing complete evaluation artifacts including metrics, executable code, and reports. To systematically assess generated evaluations, we introduce a meta-evaluation framework alongside AgentEvalBench, a benchmark comprising 20 agents, each paired with evaluation requirements and test scenarios. We further propose the Eval@1 metric to measure whether generated evaluation code both executes and yields meaningful results on the first run. Our experiments show that EvalAgent produces focused evaluations, improving Eval@1 from 17.5% to 65%, and achieving 79.5% human expert preference over baseline approaches. Further ablation studies show that evaluation skills are critical for handling complex evaluation: removing them causes Eval@1 to drop significantly from 65% to 30%.

Large Language Models (LLMs) are being used more and more extensively for automated evaluation in various scenarios. Previous studies have attempted to fine-tune open-source LLMs to replicate the evaluation explanations and judgments of powerful proprietary models, such as GPT-4. However, these methods are largely limited to text-based analyses under predefined general criteria, resulting in reduced adaptability for unseen instructions and demonstrating instability in evaluating adherence to quantitative and structural constraints. To address these limitations, we propose a novel evaluation framework, ARJudge, that adaptively formulates evaluation criteria and synthesizes both text-based and code-driven analyses to evaluate LLM responses. ARJudge consists of two components: a fine-tuned Analyzer that generates multi-faceted evaluation analyses and a tuning-free Refiner that combines and refines all analyses to make the final judgment. We construct a Composite Analysis Corpus that integrates tasks for evaluation criteria generation alongside text-based and code-driven analysis generation to train the Analyzer. Our results demonstrate that ARJudge outperforms existing fine-tuned evaluators in effectiveness and robustness. Furthermore, it demonstrates the importance of multi-faceted evaluation and code-driven analyses in enhancing evaluation capabilities.

Safety alignment in large language models relies predominantly on English-language training data. When harmful intent is expressed in low-resource languages, refusal mechanisms that hold in English frequently fail to activate. We introduce LSR (Linguistic Safety Robustness), the first systematic benchmark for measuring cross-lingual refusal degradation in West African languages: Yoruba, Hausa, Igbo, and Igala. LSR uses a dual-probe evaluation protocol - submitting matched English and target-language probes to the same model - and introduces Refusal Centroid Drift (RCD), a metric that quantifies how much of a model's English refusal behavior is lost when harmful intent is encoded in a target language. We evaluate Gemini 2.5 Flash across 14 culturally grounded attack probes in four harm categories. English refusal rates hold at approximately 90 percent. Across West African languages, refusal rates fall to 35-55 percent, with Igala showing the most severe degradation (RCD = 0.55). LSR is implemented in the Inspect AI evaluation framework and is available as a PR-ready contribution to the UK AISI's inspect_evals repository. A live reference implementation and the benchmark dataset are publicly available.

Large language models (LLMs) have achieved remarkable progress in code understanding tasks. However, they demonstrate limited performance in vulnerability detection and struggle to distinguish vulnerable code from patched code. We argue that LLMs lack understanding of security specifications -- the expectations about how code should behave to remain safe. When code behavior differs from these expectations, it becomes a potential vulnerability. However, such knowledge is rarely explicit in training data, leaving models unable to reason about security flaws. We propose VulInstruct, a specification-guided approach that systematically extracts security specifications from historical vulnerabilities to detect new ones. VulInstruct constructs a specification knowledge base from two perspectives: (i) General specifications from high-quality patches across projects, capturing fundamental safe behaviors; and (ii) Domain-specific specifications from repeated violations in particular repositories relevant to the target code. VulInstruct retrieves relevant past cases and specifications, enabling LLMs to reason about expected safe behaviors rather than relying on surface patterns. We evaluate VulInstruct under strict criteria requiring both correct predictions and valid reasoning. On PrimeVul, VulInstruct achieves 45.0% F1-score (32.7% improvement) and 37.7% recall (50.8% improvement) compared to baselines, while uniquely detecting 24.3% of vulnerabilities -- 2.4x more than any baseline. In pair-wise evaluation, VulInstruct achieves 32.3% relative improvement. VulInstruct also discovered a previously unknown high-severity vulnerability (CVE-2025-56538) in production code, demonstrating practical value for real-world vulnerability discovery. All code and supplementary materials are available at https://github.com/zhuhaopku/VulInstruct-temp.

Coding agents can modify and test code across large software projects. Game development is a domain where agents must implement gameplay rules. A game can end in a valid state even after violating its rules during the run. Current game-development benchmarks replay fixed examples, score videos, or ask another model to judge the result. However, no existing benchmark checks game rules throughout execution across varied evaluator-selected scenarios while ensuring exactly reproducible verdicts. We introduce GameLogicBench, a benchmark of 72 gameplay-logic tasks in Godot projects. An automated evaluator checks each game's rules at every simulation tick. Across 403 hand-designed scenarios, seeded parameter variations produce 1,451 test cases. To ensure that the evaluator measures behavior rather than implementation choice, it must accept different correct implementations for each task while rejecting mutants, implementations with one required capability removed. The tasks span isolated mechanics, multi-system interactions, and repository-scale features. Across 20 combinations of language models and scaffolds, the best observed run solves 52.78% of tasks. Under Claude Code, all twelve models solve fewer tasks as task scope expands from isolated mechanics, through interacting systems, to repository-scale features. Agents inspect code more often and make more tool calls on repository-scale tasks than on isolated-mechanic tasks. Most unsuccessful submissions are runnable, but implement some required game behavior incorrectly. We compared versions of our benchmark evaluator built with and without validation using mutants. Without this validation, incorrect agent submissions passed. A separate analysis finds agents copying code from public repositories when network access is open. Reliable evaluation thus depends both on what the tests reject and on what external code agents can access.

This paper provides a systematic analysis of the opportunities, challenges, and potential solutions of harnessing Large Language Models (LLMs) such as GPT-4 to dig out vulnerabilities within smart contracts based on our ongoing research. For the task of smart contract vulnerability detection, achieving practical usability hinges on identifying as many true vulnerabilities as possible while minimizing the number of false positives. Nonetheless, our empirical study reveals contradictory yet interesting findings: generating more answers with higher randomness largely boosts the likelihood of producing a correct answer but inevitably leads to a higher number of false positives. To mitigate this tension, we propose an adversarial framework dubbed GPTLens that breaks the conventional one-stage detection into two synergistic stages - generation and discrimination, for progressive detection and refinement, wherein the LLM plays dual roles, i.e., auditor and critic, respectively. The goal of auditor is to yield a broad spectrum of vulnerabilities with the hope of encompassing the correct answer, whereas the goal of critic that evaluates the validity of identified vulnerabilities is to minimize the number of false positives. Experimental results and illustrative examples demonstrate that auditor and critic work together harmoniously to yield pronounced improvements over the conventional one-stage detection. GPTLens is intuitive, strategic, and entirely LLM-driven without relying on specialist expertise in smart contracts, showcasing its methodical generality and potential to detect a broad spectrum of vulnerabilities. Our code is available at: https://github.com/git-disl/GPTLens.

Persuasion is a powerful capability of large language models (LLMs) that both enables beneficial applications (e.g. helping people quit smoking) and raises significant risks (e.g. large-scale, targeted political manipulation). Prior work has found models possess a significant and growing persuasive capability, measured by belief changes in simulated or real users. However, these benchmarks overlook a crucial risk factor: the propensity of a model to attempt to persuade in harmful contexts. Understanding whether a model will blindly ``follow orders'' to persuade on harmful topics (e.g. glorifying joining a terrorist group) is key to understanding the efficacy of safety guardrails. Moreover, understanding if and when a model will engage in persuasive behavior in pursuit of some goal is essential to understanding the risks from agentic AI systems. We propose the Attempt to Persuade Eval (APE) benchmark, that shifts the focus from persuasion success to persuasion attempts, operationalized as a model's willingness to generate content aimed at shaping beliefs or behavior. Our evaluation framework probes frontier LLMs using a multi-turn conversational setup between simulated persuader and persuadee agents. APE explores a diverse spectrum of topics including conspiracies, controversial issues, and non-controversially harmful content. We introduce an automated evaluator model to identify willingness to persuade and measure the frequency and context of persuasive attempts. We find that many open and closed-weight models are frequently willing to attempt persuasion on harmful topics and that jailbreaking can increase willingness to engage in such behavior. Our results highlight gaps in current safety guardrails and underscore the importance of evaluating willingness to persuade as a key dimension of LLM risk. APE is available at github.com/AlignmentResearch/AttemptPersuadeEval

LLM-as-a-Judge uses a large language model (LLM) to select the best response from a set of candidates for a given question. LLM-as-a-Judge has many applications such as LLM-powered search, reinforcement learning with AI feedback (RLAIF), and tool selection. In this work, we propose JudgeDeceiver, an optimization-based prompt injection attack to LLM-as-a-Judge. JudgeDeceiver injects a carefully crafted sequence into an attacker-controlled candidate response such that LLM-as-a-Judge selects the candidate response for an attacker-chosen question no matter what other candidate responses are. Specifically, we formulate finding such sequence as an optimization problem and propose a gradient based method to approximately solve it. Our extensive evaluation shows that JudgeDeceive is highly effective, and is much more effective than existing prompt injection attacks that manually craft the injected sequences and jailbreak attacks when extended to our problem. We also show the effectiveness of JudgeDeceiver in three case studies, i.e., LLM-powered search, RLAIF, and tool selection. Moreover, we consider defenses including known-answer detection, perplexity detection, and perplexity windowed detection. Our results show these defenses are insufficient, highlighting the urgent need for developing new defense strategies. Our implementation is available at this repository: https://github.com/ShiJiawenwen/JudgeDeceiver.

Adopting human and large language models (LLM) as judges (a.k.a human- and LLM-as-a-judge) for evaluating the performance of existing LLMs has recently gained attention. Nonetheless, this approach concurrently introduces potential biases from human and LLM judges, questioning the reliability of the evaluation results. In this paper, we propose a novel framework for investigating 5 types of biases for LLM and human judges. We curate a dataset with 142 samples referring to the revised Bloom's Taxonomy and conduct thousands of human and LLM evaluations. Results show that human and LLM judges are vulnerable to perturbations to various degrees, and that even the most cutting-edge judges possess considerable biases. We further exploit their weakness and conduct attacks on LLM judges. We hope that our work can notify the community of the vulnerability of human- and LLM-as-a-judge against perturbations, as well as the urgency of developing robust evaluation systems.

Frontier language models sometimes recognize that they are being evaluated and adjust their behavior, undermining validity of benchmark results. Yet the field studies it without a shared foundation, conflating properties of the evaluation with properties of the model, and detection with behavioral response. We ground evaluation awareness in social psychology, decomposing it into an environment component (how recognizable the task is) and a model component that separates recognition from propensity to act on it. We operationalize the environment component through eight categorized trigger factors, such as placeholder entities and grading-style output formats, and study recognition and behavior through chain-of-thought monitoring. Across nine frontier models and four benchmarks, recognition rates depend on the specific pairing of model and benchmark rather than on either in isolation. Recognition rarely leads to behavioral change, and when it does, the direction depends on the type of evaluation perceived. Models are also more sensitive to safety than capability evaluations, placing safety benchmark validity at greater risk. To study which factors each model is sensitive to and how they interact, we propose EvalAwareBench, a factor-controlled benchmark of 100 paired safety-capability tasks where each of the eight factors can be independently toggled, varying evaluative signals while holding the underlying request fixed. Through EvalAwareBench, we find that no single factor uniformly affects all models, but stacking factors progressively raises evaluation awareness across all of them. Our framework and EvalAwareBench provide the tools to measure, attribute, and mitigate evaluation awareness, pointing to behavioral consistency under recognition as a promising path forward.

Finetuning specialized generative evaluators has emerged as a popular paradigm to meet the increasing demand for scalable evaluation during both training and test-time. However, recent work has largely focused on applying new methodology, such as reinforcement learning (RL), to training evaluators, shying away from large-scale, data-driven development. In this work, we focus on data scaling, curating a set of 2.5M samples spanning five unique evaluation tasks (pairwise, step-level, reference-free and reference-based verification, and single rating) and multiple domains focused on reasoning evaluation. With our data, we train Foundational Automatic Reasoning Evaluators (FARE), a family of 8B and 20B (with 3.6B active) parameter evaluators, with a simple iterative rejection-sampling supervised finetuning (SFT) approach. FARE-8B challenges larger specialized RL-trained evaluators and FARE-20B sets the new standard for open-source evaluators, surpassing specialized 70B+ evaluators. Beyond static benchmarks, we evaluate FARE in real-world tasks: As inference-time rerankers, FARE-20B achieves near-oracle performance on MATH. As verifiers in RL training, FARE improves the downstream RL-trained model performance by up to 14.1% vs. string-matching verifiers. When initialized from FARE, a continually-finetuned FARE-Code outperforms gpt-oss-20B by 65% on evaluating test-case quality.

Evaluation and alignment pipelines for large language models increasingly rely on LLM-based judges, whose behavior is guided by natural-language rubrics and validated on benchmarks. We identify a previously under-recognized vulnerability in this workflow, which we term Rubric-Induced Preference Drift (RIPD). Even when rubric edits pass benchmark validation, they can still produce systematic and directional shifts in a judge's preferences on target domains. Because rubrics serve as a high-level decision interface, such drift can emerge from seemingly natural, criterion-preserving edits and remain difficult to detect through aggregate benchmark metrics or limited spot-checking. We further show this vulnerability can be exploited through rubric-based preference attacks, in which benchmark-compliant rubric edits steer judgments away from a fixed human or trusted reference on target domains, systematically inducing RIPD and reducing target-domain accuracy up to 9.5% (helpfulness) and 27.9% (harmlessness). When these judgments are used to generate preference labels for downstream post-training, the induced bias propagates through alignment pipelines and becomes internalized in trained policies. This leads to persistent and systematic drift in model behavior. Overall, our findings highlight evaluation rubrics as a sensitive and manipulable control interface, revealing a system-level alignment risk that extends beyond evaluator reliability alone. The code is available at: https://github.com/ZDCSlab/Rubrics-as-an-Attack-Surface. Warning: Certain sections may contain potentially harmful content that may not be appropriate for all readers.

Recent advances in Language Model (LM) agents and tool use, exemplified by applications like ChatGPT Plugins, enable a rich set of capabilities but also amplify potential risks - such as leaking private data or causing financial losses. Identifying these risks is labor-intensive, necessitating implementing the tools, manually setting up the environment for each test scenario, and finding risky cases. As tools and agents become more complex, the high cost of testing these agents will make it increasingly difficult to find high-stakes, long-tailed risks. To address these challenges, we introduce ToolEmu: a framework that uses an LM to emulate tool execution and enables the testing of LM agents against a diverse range of tools and scenarios, without manual instantiation. Alongside the emulator, we develop an LM-based automatic safety evaluator that examines agent failures and quantifies associated risks. We test both the tool emulator and evaluator through human evaluation and find that 68.8% of failures identified with ToolEmu would be valid real-world agent failures. Using our curated initial benchmark consisting of 36 high-stakes tools and 144 test cases, we provide a quantitative risk analysis of current LM agents and identify numerous failures with potentially severe outcomes. Notably, even the safest LM agent exhibits such failures 23.9% of the time according to our evaluator, underscoring the need to develop safer LM agents for real-world deployment.

Automated evaluation of tool-using large language model (LLM) agents is widely assumed to be reliable, but this assumption has rarely been validated against human annotation. We introduce AgentProp-Bench, a 2,000-task benchmark with 2,300 traces across four domains, nine production LLMs, and a 100-label human-validated subset. We quantify judge reliability, characterize error propagation, and evaluate a runtime mitigation. Substring-based judging agrees with human annotation at kappa=0.049 (chance-level); a three-LLM ensemble reaches kappa=0.432 (moderate) with a conservative bias. Under validated evaluation, a parameter-level injection propagates to a wrong final answer with human-calibrated probability approximately 0.62 (range 0.46-0.73 across models). Rejection (catching bad parameters) and recovery (correcting after acceptance) are independent model capabilities (Spearman rho=0.126, p=0.747). A tuned runtime interceptor reduces hallucination on GPT-4o-mini by 23.0 percentage points under a concurrent n=600 control, but shows no significant effect on Gemini-2.0-Flash, whose aggressive parameter rejection eliminates the target failure mode. All code, data, traces, and human labels are released at https://github.com/bhaskargurram-ai/agenthallu-bench.

The rapid deployment of Large Language Models (LLMs) requires careful consideration of their effect on cybersecurity. Our work aims to improve the selection process of LLMs that are suitable for facilitating Secure Coding (SC). This raises challenging research questions, such as (RQ1) Which functionality can streamline the LLM evaluation? (RQ2) What should the evaluation measure? (RQ3) How to attest that the evaluation process is impartial? To address these questions, we introduce LLMSecCode, an open-source evaluation framework designed to assess LLM SC capabilities objectively. We validate the LLMSecCode implementation through experiments. When varying parameters and prompts, we find a 10% and 9% difference in performance, respectively. We also compare some results to reliable external actors, where our results show a 5% difference. We strive to ensure the ease of use of our open-source framework and encourage further development by external actors. With LLMSecCode, we hope to encourage the standardization and benchmarking of LLMs' capabilities in security-oriented code and tasks.

Large language models (LLMs) have shown promising performance in software vulnerability detection, yet their reasoning capabilities remain unreliable. We propose R2Vul, a method that combines reinforcement learning from AI feedback (RLAIF) and structured reasoning distillation to teach small code LLMs to detect vulnerabilities while generating security-aware explanations. Unlike prior chain-of-thought and instruction tuning approaches, R2Vul rewards well-founded over deceptively plausible vulnerability explanations through RLAIF, which results in more precise detection and high-quality reasoning generation. To support RLAIF, we construct the first multilingual preference dataset for vulnerability detection, comprising 18,000 high-quality samples in C\#, JavaScript, Java, Python, and C. We evaluate R2Vul across five programming languages and against four static analysis tools, eight state-of-the-art LLM-based baselines, and various fine-tuning approaches. Our results demonstrate that a 1.5B R2Vul model exceeds the performance of its 32B teacher model and leading commercial LLMs such as Claude-4-Opus. Furthermore, we introduce a lightweight calibration step that reduces false positive rates under varying imbalanced data distributions. Finally, through qualitative analysis, we show that both LLM and human evaluators consistently rank R2Vul model's reasoning higher than other reasoning-based baselines.

Agent benchmarks typically report only final outcomes: pass or fail. This threatens evaluation credibility in three ways. First, scores may be inflated or deflated by shortcuts and benchmark artifacts, misrepresenting capability. Second, benchmark performance may fail to predict real-world utility due to scaffold limitations and recurring failure modes. Finally, capability scores may conceal dangerous or catastrophic actions taken by the agent. We argue that log analysis -- the systematic tracking and analysis of the inputs, execution, and outputs of an AI agent -- is necessary to overcome these validity threats and promote credible agent evaluation. In this paper, we (1) present a taxonomy of threats to credible evaluation documented through log analysis, and (2) develop a set of guiding principles for log analysis. We illustrate these principles on tau-Bench Airline, revealing that pass^5 performance was under-elicited by nearly 50% and surfacing deployment failure modes invisible to outcome metrics. We conclude with pragmatic recommendations to increase uptake of log analysis, directed at diverse stakeholders including benchmark creators, model developers, independent evaluators, and deployers.

Deploying language models (LMs) necessitates outputs to be both high-quality and compliant with safety guidelines. Although Inference-Time Guardrails (ITG) offer solutions that shift model output distributions towards compliance, we find that current methods struggle in balancing safety with helpfulness. ITG Methods that safely address non-compliant queries exhibit lower helpfulness while those that prioritize helpfulness compromise on safety. We refer to this trade-off as the guardrail tax, analogous to the alignment tax. To address this, we propose PrimeGuard, a novel ITG method that utilizes structured control flow. PrimeGuard routes requests to different self-instantiations of the LM with varying instructions, leveraging its inherent instruction-following capabilities and in-context learning. Our tuning-free approach dynamically compiles system-designer guidelines for each query. We construct and release safe-eval, a diverse red-team safety benchmark. Extensive evaluations demonstrate that PrimeGuard, without fine-tuning, overcomes the guardrail tax by (1) significantly increasing resistance to iterative jailbreak attacks and (2) achieving state-of-the-art results in safety guardrailing while (3) matching helpfulness scores of alignment-tuned models. Extensive evaluations demonstrate that PrimeGuard, without fine-tuning, outperforms all competing baselines and overcomes the guardrail tax by improving the fraction of safe responses from 61% to 97% and increasing average helpfulness scores from 4.17 to 4.29 on the largest models, while reducing attack success rate from 100% to 8%. PrimeGuard implementation is available at https://github.com/dynamofl/PrimeGuard and safe-eval dataset is available at https://huggingface.co/datasets/dynamoai/safe_eval.

Memory safety vulnerabilities remain a significant threat even for projects with extensive fuzzing and manual auditing. Recent results suggest that large language models hold great promise for detecting such vulnerabilities, but they are unreliable, at risk of hallucination, and challenging to scale to repository-size codebases. This paper presents Revelio, a cost-efficient end-to-end agentic framework for memory-safety vulnerability discovery. Revelio addresses the problem of hallucination by generating an executable Proof-of-Vulnerability, which is checked with a deterministic sanitizer. It reduces cost using inexpensive LLMs and lightweight static analysis to help generate and rank vulnerability hypotheses, reporting vulnerabilities only when they can be reproduced and confirmed by a sanitizer. We evaluated Revelio on seven production-quality projects that had been continuously fuzzed for five to eight years, as well as on 100 randomly selected Arvo projects from the CyberGym benchmark. With around one hour per project and a total cost of $300, Revelio discovered 19 previously unknown memory-safety vulnerabilities. On benchmarks, Revelio outperformed frontier coding agents across diverse backbone models at comparable token costs. Our results suggest that Revelio enables scalable and trustworthy end-to-end LLM-based memory-safety vulnerability detection.

Reliable evaluation is essential for developing and deploying large language models, yet in practice it often requires substantial manual effort: practitioners must identify appropriate benchmarks, reproduce heterogeneous evaluation codebases, configure dataset schema mappings, and interpret aggregated metrics. To address these challenges, we present One-Eval, an agentic evaluation system that converts natural-language evaluation requests into executable, traceable, and customizable evaluation workflows. One-Eval integrates (i) NL2Bench for intent structuring and personalized benchmark planning, (ii) BenchResolve for benchmark resolution, automatic dataset acquisition, and schema normalization to ensure executability, and (iii) Metrics \& Reporting for task-aware metric selection and decision-oriented reporting beyond scalar scores. The system further incorporates human-in-the-loop checkpoints for review, editing, and rollback, while preserving sample evidence trails for debugging and auditability. Experiments show that One-Eval can execute end-to-end evaluations from diverse natural-language requests with minimal user effort, supporting more efficient and reproducible evaluation in industrial settings. Our framework is publicly available at https://github.com/OpenDCAI/One-Eval.

Despite extensive safety alignment, large language models remain vulnerable to jailbreak attacks. Existing optimization-based attacks share a single objective: maximize a token-level proxy of victim compliance---such as the likelihood of an affirmative prefix---computed on the prompt rather than the decoded reply. This proxy cannot distinguish a cautious hedge from a fully actionable harmful answer, leaving the search blind to what the model actually says. We propose ECHO (Evaluation-Conditioned Harmfulness Optimization), which replaces the compliance proxy with response-level, continuous harmfulness scores from HarmSig, a 1.7B regressor grounded in actual harmful capability. ECHO couples HarmSig with five word-level perturbation operators and Keyword-Anchored Beam Search with global top-m memory, under a strict black-box threat model. Across five open-weight victims on AdvBench, ECHO achieves up to 0.898 StrongREJECT mean score and 89.23% HarmBench success. Our code is publicly available at https://github.com/STANWYW/echo.

Large language model (LLM) agents combine language-based reasoning with external tools to perform complex tasks. Adversarial inputs can exploit interactions between the agent and its environment, causing the agent to violate safety policies during execution. Yet existing evaluations often reduce agent safety to a single attack success rate (ASR), collapsing exposure, execution, observation, and adjudication and potentially conflating actual violations with evidence visibility. We introduce REDAgentBench, an executable framework for autonomous red-teaming and faithful measurement. It derives attacks from explicit safety constraints and associated agent-system vulnerabilities, runs them in isolated service sandboxes, and verifies harmful effects from service receipts and final-state changes. The benchmark contains 1,661 cases across five service surfaces. Across six models and three agent harnesses, macro-average ASR is 65.69%; reported ASR varies with harness and evidence view, while evaluation-context disclosure changes execution behavior. In a state-grounded diagnostic cohort, almost one in five confirmed violations with resolved action anchors occurs after the agent states the relevant constraint or risk, revealing a Recognition--Execution Gap. Finally, a training-free policy reminder reduces confirmed violations by more than 70 percentage points in matched replay. These findings show that executable evaluation can improve safety measurement and identify actionable intervention points.

Large Reasoning Models (LRMs) have recently extended their powerful reasoning capabilities to safety checks-using chain-of-thought reasoning to decide whether a request should be answered. While this new approach offers a promising route for balancing model utility and safety, its robustness remains underexplored. To address this gap, we introduce Malicious-Educator, a benchmark that disguises extremely dangerous or malicious requests beneath seemingly legitimate educational prompts. Our experiments reveal severe security flaws in popular commercial-grade LRMs, including OpenAI o1/o3, DeepSeek-R1, and Gemini 2.0 Flash Thinking. For instance, although OpenAI's o1 model initially maintains a high refusal rate of about 98%, subsequent model updates significantly compromise its safety; and attackers can easily extract criminal strategies from DeepSeek-R1 and Gemini 2.0 Flash Thinking without any additional tricks. To further highlight these vulnerabilities, we propose Hijacking Chain-of-Thought (H-CoT), a universal and transferable attack method that leverages the model's own displayed intermediate reasoning to jailbreak its safety reasoning mechanism. Under H-CoT, refusal rates sharply decline-dropping from 98% to below 2%-and, in some instances, even transform initially cautious tones into ones that are willing to provide harmful content. We hope these findings underscore the urgent need for more robust safety mechanisms to preserve the benefits of advanced reasoning capabilities without compromising ethical standards.

Manual curation of evaluation datasets is struggling to keep up with the rapidly expanding capabilities and deployment scenarios of language models. Towards scalable model profiling, we introduce and validate a framework for evaluating LLMs, called Adaptive Evaluations. Adaptive evaluations use scaffolded language models (evaluator agents) to search through a target model's behavior on a domain dataset and create difficult questions (tasks) that can discover and probe the model's failure modes. We find that frontier models lack consistency when adaptively probed with our framework on a diverse suite of datasets and tasks, including but not limited to legal reasoning, forecasting, and online harassment. Generated questions pass human validity checks and often transfer to other models with different capability profiles, demonstrating that adaptive evaluations can also be used to create difficult domain-specific datasets.

The code generation capabilities of large language models(LLMs) have emerged as a critical dimension in evaluating their overall performance. However, prior research has largely overlooked the security risks inherent in the generated code. In this work, we introduce SafeGenBench, a benchmark specifically designed to assess the security of LLM-generated code. The dataset encompasses a wide range of common software development scenarios and vulnerability types. Building upon this benchmark, we develop an automatic evaluation framework that leverages both static application security testing(SAST) and LLM-based judging to assess the presence of security vulnerabilities in model-generated code. Through the empirical evaluation of state-of-the-art LLMs on SafeGenBench, we reveal notable deficiencies in their ability to produce vulnerability-free code. Our findings highlight pressing challenges and offer actionable insights for future advancements in the secure code generation performance of LLMs. The data and code will be released soon.

Executable evaluation -- checking the consequences of an agent's actions with a program rather than grading its prose -- has become a prominent way to assess tool-using AI agents in software settings. Electric power engineering has not yet had an analogous benchmark: language-model use is still dominated by retrieval and text question answering, while agents acting on power-system artifacts remain mostly academic prototypes. We introduce the Power Systems Agent Benchmark, an executable benchmark for power-engineering agents. An agent receives a structured task and returns a structured solution; a deterministic evaluator recomputes the engineering quantities, checks operational constraints, and returns a feasibility flag, a normalized score, and explicit violations. The benchmark contains 41 task families across eight areas of power engineering, from power flow and protection to stability, microgrids, reliability, power quality, and forecasting. Each task is grounded in a citable source, standard, or documented engineering formulation. To resist contamination, held-out cases are synthesized on demand by per-family generators from private seeds: the construction is inspectable, but the instances remain private. In a reference evaluation with three command-line agents, the strongest score near the compact tier's ceiling, a smaller open model trails, and public and held-out performance are broadly consistent; a separate public-split grid with OpenCode and Aider probes harness effects. The reference evaluation doubles as quality control: unanimous failures flag candidate task or evaluator defects, and it exposed a latent evaluator bug missed by self-consistency checks. The evaluators are compact deterministic surrogates, but the task contract allows their internals to be upgraded to simulator-backed checks without changing how tasks are posed or solved.

Current learning-based Automated Vulnerability Repair (AVR) approaches, while promising, often fail to generalize effectively in real-world scenarios. Our diagnostic analysis reveals three fundamental weaknesses in state-of-the-art AVR approaches: (1) limited cross-repository generalization, with performance drops on unseen codebases; (2) inability to capture long-range dependencies, causing a performance degradation on complex, multi-hunk repairs; and (3) over-reliance on superficial lexical patterns, leading to significant performance drops on vulnerabilities with minor syntactic variations like variable renaming. To address these limitations, we propose SeCuRepair, a semantics-aligned, curriculum-driven, and reasoning-enhanced framework for vulnerability repair. At its core, SeCuRepair adopts a reason-then-edit paradigm, requiring the model to articulate why and how a vulnerability should be fixed before generating the patch. This explicit reasoning enforces a genuine understanding of repair logic rather than superficial memorization of lexical patterns. SeCuRepair also moves beyond traditional supervised fine-tuning and employs semantics-aware reinforcement learning, rewarding patches for their syntactic and semantic alignment with the oracle patch rather than mere token overlap. Complementing this, a difficulty-aware curriculum progressively trains the model, starting with simple fixes and advancing to complex, multi-hunk coordinated edits. We evaluate SeCuRepair on strict, repository-level splits of BigVul and newly crafted PrimeVul_AVR datasets. SeCuRepair significantly outperforms all baselines, surpassing the best-performing baselines by 34.52% on BigVul and 31.52% on PrimeVulAVR in terms of CodeBLEU, respectively. Comprehensive ablation studies further confirm that each component of our framework contributes to its final performance.

Model Inversion (MI) attacks aim to reconstruct information from private training data by exploiting access to machine learning models T. To evaluate such attacks, the standard evaluation framework relies on an evaluation model E, trained under the same task design as T. This framework has become the de facto standard for assessing progress in MI research, used across nearly all recent MI studies without question. In this paper, we present the first in-depth study of this evaluation framework. In particular, we identify a critical issue of this standard framework: Type-I adversarial examples. These are reconstructions that do not capture the visual features of private training data, yet are still deemed successful by T and ultimately transferable to E. Such false positives undermine the reliability of the standard MI evaluation framework. To address this issue, we introduce a new MI evaluation framework that replaces the evaluation model E with advanced Multimodal Large Language Models (MLLMs). By leveraging their general-purpose visual understanding, our MLLM-based framework does not depend on training of shared task design as in T, thus reducing Type-I transferability and providing more faithful assessments of reconstruction success. Using our MLLM-based evaluation framework, we reevaluate 27 diverse MI attack setups and empirically reveal consistently high false positive rates under the standard evaluation framework. Importantly, we demonstrate that many state-of-the-art (SOTA) MI methods report inflated attack accuracy, indicating that actual privacy leakage is significantly lower than previously believed. By uncovering this critical issue and proposing a robust solution, our work enables a reassessment of progress in MI research and sets a new standard for reliable and robust evaluation. Code can be found in https://github.com/hosytuyen/MI-Eval-MLLM

LLM agents often retry after external validation rejects a candidate, but the interface between validation and the next model call remains underspecified. We introduce VeriHarness, a code-controlled agent loop in which models generate candidates while external validators control acceptance, budgets, and traces. We use it to compare raw diagnostics with feedback that identifies the failure location, observed value, and admissible alternatives. Across 50 paired TextWorld games under a four-call cap, feedback containing all three fields raises terminal success from 14/50 to 36/50 for Qwen2.5-Coder-14B (+44 percentage points) and from 8/50 to 29/50 for Llama-3.1-8B (+42 points). Ablations locate most of the gain in the admissible alternatives: feedback containing only the location and observed value remains near the raw diagnostic baseline. Presenting the complete repair information in prose instead of a keyed JSON record yields nearly the same success, providing no evidence that JSON syntax itself improves repair. The ordering persists across the tested call budgets and one sampled-decoding setting.

Software vulnerabilities are increasing at an alarming rate. However, manual patching is both time-consuming and resource-intensive, while existing automated vulnerability repair (AVR) techniques remain limited in effectiveness. Recent advances in large language models (LLMs) have opened a new paradigm for AVR, demonstrating remarkable progress. To examine the capability of LLMs in AVR, several vulnerability benchmarks have been proposed recently. However, they still suffer from key limitations of outdated vulnerabilities, limited language coverage, unreliable patch validation, and insufficient reproducibility. To overcome these challenges, we introduce PATCHEVAL, a multilingual benchmark for Go, JavaScript, and Python, languages for which existing benchmarks remain unexplored. PATCHEVAL curates a dataset of 1,000 vulnerabilities drawn from CVEs reported between 2015 and 2025, covering 65 distinct CWEs. A subset of 230 CVEs is further equipped with runtime sandbox environments, enabling patch verification through both security tests and functionality tests. To provide a systematic comparison of LLM-based vulnerability repair, we evaluate a series of state-of-the-art LLMs and agents, presenting an in-depth analysis that empirically yields key insights to guide future research in AVR.

Large Language Models (LLMs) are powerful zero-shot assessors and are increasingly used in real-world situations such as for written exams or benchmarking systems. Despite this, no existing work has analyzed the vulnerability of judge-LLMs against adversaries attempting to manipulate outputs. This work presents the first study on the adversarial robustness of assessment LLMs, where we search for short universal phrases that when appended to texts can deceive LLMs to provide high assessment scores. Experiments on SummEval and TopicalChat demonstrate that both LLM-scoring and pairwise LLM-comparative assessment are vulnerable to simple concatenation attacks, where in particular LLM-scoring is very susceptible and can yield maximum assessment scores irrespective of the input text quality. Interestingly, such attacks are transferable and phrases learned on smaller open-source LLMs can be applied to larger closed-source models, such as GPT3.5. This highlights the pervasive nature of the adversarial vulnerabilities across different judge-LLM sizes, families and methods. Our findings raise significant concerns on the reliability of LLMs-as-a-judge methods, and underscore the importance of addressing vulnerabilities in LLM assessment methods before deployment in high-stakes real-world scenarios.

Recognizing vulnerability is crucial for understanding and implementing targeted support to empower individuals in need. This is especially important at the European Court of Human Rights (ECtHR), where the court adapts Convention standards to meet actual individual needs and thus ensures effective human rights protection. However, the concept of vulnerability remains elusive at the ECtHR and no prior NLP research has dealt with it. To enable future research in this area, we present VECHR, a novel expert-annotated multi-label dataset comprising of vulnerability type classification and explanation rationale. We benchmark the performance of state-of-the-art models on VECHR from both prediction and explainability perspectives. Our results demonstrate the challenging nature of the task with lower prediction performance and limited agreement between models and experts. Further, we analyze the robustness of these models in dealing with out-of-domain (OOD) data and observe overall limited performance. Our dataset poses unique challenges offering significant room for improvement regarding performance, explainability, and robustness.

Long contexts of recent LLMs have enabled a new use case: asking models to find security vulnerabilities in entire codebases. To evaluate model performance on this task, we introduce eyeballvul: a benchmark designed to test the vulnerability detection capabilities of language models at scale, that is sourced and updated weekly from the stream of published vulnerabilities in open-source repositories. The benchmark consists of a list of revisions in different repositories, each associated with the list of known vulnerabilities present at that revision. An LLM-based scorer is used to compare the list of possible vulnerabilities returned by a model to the list of known vulnerabilities for each revision. As of July 2024, eyeballvul contains 24,000+ vulnerabilities across 6,000+ revisions and 5,000+ repositories, and is around 55GB in size.

This technical report accompanies the LLM robustness leaderboard published by PRISM Eval for the Paris AI Action Summit. We introduce PRISM Eval Behavior Elicitation Tool (BET), an AI system performing automated red-teaming through Dynamic Adversarial Optimization that achieves 100% Attack Success Rate (ASR) against 37 of 41 state-of-the-art LLMs. Beyond binary success metrics, we propose a fine-grained robustness metric estimating the average number of attempts required to elicit harmful behaviors, revealing that attack difficulty varies by over 300-fold across models despite universal vulnerability. We introduce primitive-level vulnerability analysis to identify which jailbreaking techniques are most effective for specific hazard categories. Our collaborative evaluation with trusted third parties from the AI Safety Network demonstrates practical pathways for distributed robustness assessment across the community.

Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Moreover, they rarely evaluate a model's ability to detect culturally embedded sensitivities as distinct from universal harms. We introduce XL-SafetyBench. a suite of 5,500 test cases across 10 country-language pairs, comprising a Jailbreak Benchmark of country-grounded adversarial prompts and a Cultural Benchmark where local sensitivities are embedded within innocuous requests. Each item is constructed via a multi-stage pipeline that combines LLM-assisted discovery, automated validation gates, and dual independent native-speaker annotators per country. To distinguish principled refusal from comprehension failure, we evaluate Attack Success Rate (ASR) alongside two complementary metrics we introduce: Neutral-Safe Rate (NSR) and Cultural Sensitivity Rate (CSR). Evaluating 10 frontier and 27 local LLMs reveals two key findings. First, jailbreak robustness and cultural awareness do not show a coupled relationship among frontier models, so a composite safety score obscures per-axis variation. Second, local models exhibit a near-linear ASR-NSR trade-off (r = -0.81), indicating that their apparent safety reflects generation failure rather than genuine alignment. XL-SafetyBench enables more nuanced, cross-cultural safety evaluation in the multilingual era.

Evaluating fairness can be challenging in practice because the sensitive attributes of data are often inaccessible due to privacy constraints. The go-to approach that the industry frequently adopts is using off-the-shelf proxy models to predict the missing sensitive attributes, e.g. Meta [Alao et al., 2021] and Twitter [Belli et al., 2022]. Despite its popularity, there are three important questions unanswered: (1) Is directly using proxies efficacious in measuring fairness? (2) If not, is it possible to accurately evaluate fairness using proxies only? (3) Given the ethical controversy over inferring user private information, is it possible to only use weak (i.e. inaccurate) proxies in order to protect privacy? Our theoretical analyses show that directly using proxy models can give a false sense of (un)fairness. Second, we develop an algorithm that is able to measure fairness (provably) accurately with only three properly identified proxies. Third, we show that our algorithm allows the use of only weak proxies (e.g. with only 68.85%accuracy on COMPAS), adding an extra layer of protection on user privacy. Experiments validate our theoretical analyses and show our algorithm can effectively measure and mitigate bias. Our results imply a set of practical guidelines for practitioners on how to use proxies properly. Code is available at github.com/UCSC-REAL/fair-eval.

Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component (η^2 approx 0.52), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.

When a phone-use agent avoids harm, does that show safety, or simply inability to act? Existing evaluations often cannot tell. A harmful outcome may be avoided because the agent recognized the risk and chose the safe action, or because it failed to understand the screen or execute any relevant action at all. These cases have different causes and call for different fixes, yet current benchmarks often merge them under task success, refusal, or final harmful outcome. We address this problem with PhoneSafety, a benchmark of 700 safety-critical moments drawn from real phone interactions across more than 130 apps. Each instance isolates the next decision at a risky moment and asks a simple question: does the model take the safe action, take the unsafe action, or fail to do anything useful? We evaluate eight representative phone-use agents under this framework. Our results reveal two main patterns. First, stronger general phone-use ability does not reliably imply safer choices at risky moments. Models that perform better on ordinary app tasks are not always the ones that behave more safely when the next action matters. Second, failures to do anything useful behave like a capability signal rather than a safety signal: they are concentrated in more visually and operationally demanding settings and remain stable when the evaluation protocol changes. Across models, failures split into two recurring patterns: unsafe choices in settings where the model can act but chooses wrongly, and inability to act in more visually and operationally demanding screens. Overall, a harmless outcome is not enough to count as evidence of safety. Evaluating phone-use agents requires separating unsafe judgment from inability to act.

Large Language Model (LLM) safety has often been evaluated at the behavior level, which provides limited evidence of internal robustness, as these evaluations target outputs rather than representation-level vulnerability under intervention. We formalize this discrepancy as the audit gap: the difference between behavioral safety and robustness under intervention. To study this gap, we construct dissociated models that preserve safe outward behavior while remaining vulnerable in the latent space. We introduce an intervention-based evaluation framework to test model robustness through soft interventions in parameter and latent spaces, including harmful fine-tuning and layer-wise latent perturbations. To formalize the evaluation, we propose the Latent Vulnerability Score (LVS) to measure how easily harmful behavior can be elicited by bounded latent perturbations. Using this evaluation framework, we show that behavioral safety metrics are insufficient measures of representation-level robustness across multiple safely and unsafely aligned state-of-the-art models. Notably, dissociated models show substantially elevated LVSs despite comparable refusal behavior under harmful intervention, with intermediate representations being the most sensitive to intervention. Our results suggest that behavioral safety evaluation alone provides an incomplete picture of model robustness, motivating representation-aware audits of latent vulnerability and observable behavior.

Large language models (LLMs) are increasingly explored as scalable tools for mental health counseling, yet evaluating their safety remains challenging due to the interactional and context-dependent nature of clinical harm. Existing evaluation frameworks predominantly assess isolated responses using coarse-grained taxonomies or static datasets, limiting their ability to diagnose how harms emerge and accumulate over multi-turn counseling interactions. In this work, we introduce R-MHSafe, a role-aware mental health safety taxonomy that characterizes clinically significant harm in terms of the interactional roles an AI counselor adopts, including perpetrator, instigator, facilitator, or enabler, combined with clinically grounded harm categories. Then, we propose MHSafeEval, a closed-loop, agent-based evaluation framework that formulates safety assessment as trajectory-level discovery of harm through adversarial multi-turn interactions, guided by role-aware modeling. Using R-MHSafe and MHSafeEval, we conduct a large-scale evaluation across state-of-the-art LLMs. Our results reveal substantial role-dependent and cumulative safety failures that are systematically missed by existing static benchmarks, and show that our framework significantly improves failure-mode coverage and diagnostic granularity.

Evaluations of large language model (LLM) risks and capabilities are increasingly being incorporated into AI risk management and governance frameworks. Currently, most risk evaluations are conducted by designing inputs that elicit harmful behaviors from the system. However, a fundamental limitation of this approach is that the harmfulness of the behaviors identified during any particular evaluation can only lower bound the model's worst-possible-case behavior. As a complementary method for eliciting harmful behaviors, we propose evaluating LLMs with model tampering attacks which allow for modifications to latent activations or weights. We pit state-of-the-art techniques for removing harmful LLM capabilities against a suite of 5 input-space and 6 model tampering attacks. In addition to benchmarking these methods against each other, we show that (1) model resilience to capability elicitation attacks lies on a low-dimensional robustness subspace; (2) the attack success rate of model tampering attacks can empirically predict and offer conservative estimates for the success of held-out input-space attacks; and (3) state-of-the-art unlearning methods can easily be undone within 16 steps of fine-tuning. Together these results highlight the difficulty of removing harmful LLM capabilities and show that model tampering attacks enable substantially more rigorous evaluations than input-space attacks alone. We release models at https://huggingface.co/LLM-GAT

As large language models (LLMs) become increasingly capable, it is prudent to assess whether safety measures remain effective even if LLMs intentionally try to bypass them. Previous work introduced control evaluations, an adversarial framework for testing deployment strategies of untrusted models (i.e., models which might be trying to bypass safety measures). While prior work treats a single failure as unacceptable, we perform control evaluations in a "distributed threat setting" -- a setting where no single action is catastrophic and no single action provides overwhelming evidence of misalignment. We approach this problem with a two-level deployment framework that uses an adaptive macro-protocol to choose between micro-protocols. Micro-protocols operate on a single task, using a less capable, but extensively tested (trusted) model to harness and monitor the untrusted model. Meanwhile, the macro-protocol maintains an adaptive credence on the untrusted model's alignment based on its past actions, using it to pick between safer and riskier micro-protocols. We evaluate our method in a code generation testbed where a red team attempts to generate subtly backdoored code with an LLM whose deployment is safeguarded by a blue team. We plot Pareto frontiers of safety (# of non-backdoored solutions) and usefulness (# of correct solutions). At a given level of usefulness, our adaptive deployment strategy reduces the number of backdoors by 80% compared to non-adaptive baselines.

Automated detection of software vulnerabilities is a fundamental problem in software security. Existing program analysis techniques either suffer from high false positives or false negatives. Recent progress in Deep Learning (DL) has resulted in a surge of interest in applying DL for automated vulnerability detection. Several recent studies have demonstrated promising results achieving an accuracy of up to 95% at detecting vulnerabilities. In this paper, we ask, "how well do the state-of-the-art DL-based techniques perform in a real-world vulnerability prediction scenario?". To our surprise, we find that their performance drops by more than 50%. A systematic investigation of what causes such precipitous performance drop reveals that existing DL-based vulnerability prediction approaches suffer from challenges with the training data (e.g., data duplication, unrealistic distribution of vulnerable classes, etc.) and with the model choices (e.g., simple token-based models). As a result, these approaches often do not learn features related to the actual cause of the vulnerabilities. Instead, they learn unrelated artifacts from the dataset (e.g., specific variable/function names, etc.). Leveraging these empirical findings, we demonstrate how a more principled approach to data collection and model design, based on realistic settings of vulnerability prediction, can lead to better solutions. The resulting tools perform significantly better than the studied baseline: up to 33.57% boost in precision and 128.38% boost in recall compared to the best performing model in the literature. Overall, this paper elucidates existing DL-based vulnerability prediction systems' potential issues and draws a roadmap for future DL-based vulnerability prediction research. In that spirit, we make available all the artifacts supporting our results: https://git.io/Jf6IA.

Refusal rates are a poor proxy for LLM safety, i.e., a model may over-refuse benign prompts while still complying with harmful ones. We audit both failure modes across 21 open-weight LLMs on four safety benchmarks (OR-Bench, XSTest, ToxiGen, BOLD), using a composition adjustment to isolate model sensitivity from dataset toxicity confounds. We report three findings. First, models adopt fundamentally different calibration strategies: conservative ecosystems such as Llama suppress unsafe outputs at the cost of elevated over-refusals, while permissive ecosystems such as DeepSeek and Qwen preserve helpfulness but tolerate higher harmful compliance. Second, demographic protection is unequal: models over-protect prominent racial and religious groups, frequently refusing even benign prompts about them, while providing substantially weaker protection against disability-targeted attacks. Third, refusal and compliance tendencies are stable within model families across generations and scales, suggesting that post-training objectives shape safety behavior more than architecture. Our results call for joint, demographically-aware, and multi-judge safety evaluation.

Leading language model (LM) providers like OpenAI and Google offer fine-tuning APIs that allow customers to adapt LMs for specific use cases. To prevent misuse, these LM providers implement filtering mechanisms to block harmful fine-tuning data. Consequently, adversaries seeking to produce unsafe LMs via these APIs must craft adversarial training data that are not identifiably harmful. We make three contributions in this context: 1. We show that many existing attacks that use harmless data to create unsafe LMs rely on eliminating model refusals in the first few tokens of their responses. 2. We show that such prior attacks can be blocked by a simple defense that pre-fills the first few tokens from an aligned model before letting the fine-tuned model fill in the rest. 3. We describe a new data-poisoning attack, ``No, Of course I Can Execute'' (NOICE), which exploits an LM's formulaic refusal mechanism to elicit harmful responses. By training an LM to refuse benign requests on the basis of safety before fulfilling those requests regardless, we are able to jailbreak several open-source models and a closed-source model (GPT-4o). We show an attack success rate (ASR) of 57% against GPT-4o; our attack earned a Bug Bounty from OpenAI. Against open-source models protected by simple defenses, we improve ASRs by an average of 3.25 times compared to the best performing previous attacks that use only harmless data. NOICE demonstrates the exploitability of repetitive refusal mechanisms and broadens understanding of the threats closed-source models face from harmless data.

Instruction-tuned large language models produce helpful, structured responses, but how robust is this helpfulness under trivial constraints? We show that simple lexical constraints (banning a single punctuation character or common word) cause instruction-tuned LLMs to collapse their responses, losing 14--48\% of comprehensiveness across seven models spanning five families (7B--70B, open- and closed-weight). A blinded human evaluation with 10 STEM-trained evaluators confirms genuine content loss, with information criteria degrading 1.5--2.3times more than surface criteria, a finding corroborated by over 4,100 automated pairwise comparisons (77--100\% baseline preference) across three LLM judges from two model families. Diagnostic analysis identifies this as a planning failure: two-pass generation recovers 59--96\% of response length, and linear probes on prompt representations predict response length with R^2 = 0.51--0.94 before generation begins. The same probes yield negative R^2 on base models, confirming that instruction tuning introduces the representational structure underlying the collapse. Base models show no systematic degradation under identical constraints, demonstrating that instruction tuning couples task competence to narrow surface-form templates. The effect extends to realistic deployment constraints (preamble suppression, corporate tone guidelines, legal compliance hedging, accessibility requirements) causing comparable degradation (-22\% to -34\%), with suppressing the conversational opener alone (``Certainly!'') causing 40\% collapse on our most fragile model despite restricting only the opening tokens. We further show that standard independent LLM-as-judge evaluation detects only a 3.5\% quality drop where pairwise evaluation reveals 23\%, exposing a methodological blind spot in current evaluation practice.

Large language models are increasingly used to author Infrastructure-as-Code (IaC), where a single insecure default can be deployed directly into production. Prior evaluations report raw vulnerability counts for model-generated IaC, but without a human baseline they cannot determine whether models are actually worse than engineers. We introduce GenIaC-SecBench, a benchmark of 100 deployment scenarios stratified by architectural complexity, evaluated across 12 model configurations from four vendors, producing 1,196 IaC artifacts scanned by three independent policy engines (Checkov, Trivy, KICS). Critically, we also scan 634 human-authored IaC templates with the same toolchain, providing the first size-matched human security baseline. Vulnerability density is strongly inverse to artifact size (Spearman ρ= -0.55, p < 10^{-77}), meaning unmatched comparisons measure size rather than security. When matched on declared-resource count, all model configurations fall within 3.21x--3.87x the human vulnerability density, with the gap widening for simpler tasks (4.9x at one resource, 1.4x at twenty or more). We decompose reasoning into standard generation, prompt-engineered chain-of-thought, and vendor extended-thinking APIs. Vendor extended thinking significantly outperforms prompted chain-of-thought (-12.0%, p = 0.0013), while prompted chain-of-thought is indistinguishable from standard generation (-1.3%, n.s.). Token instrumentation shows extended thinking uses under 1\% of the output budget, explaining its bounded effect. Two negative results also emerge: deployability does not correlate with vulnerability (r = 0.158, p = 0.625), and classical complete-case Friedman testing is infeasible for realistic benchmark designs, motivating the Skillings-Mack statistic. All code, data, and regeneration scripts are released.

Agent skill libraries are becoming routable software assets: a retrieved skill can contribute instructions, scripts, resource bindings, and execution assumptions to an agent. This makes retrieval failures more specific than broad irrelevance. A system can find the right capability family yet expose the wrong same-capability representative. We study this failure as same-capability risk-exposure retrieval. Each benchmark unit pairs a helpful skill with a query-specific risky sibling that shares the capability family but differs on an execution-controlling contract, such as the required resource, precondition, procedure, or artifact. We introduce SameCapRisk-Bench, an auditable benchmark with 1,190 skill-risk units and 1,686 evaluation query cases: 694 marked-sibling units under public library pressure and 496 hard role-flip units where the same two skills swap helpful/risky roles across paired queries. The release records admission evidence, cue/leakage checks, source hashes, family relations, and fixed candidate pools. The benchmark reports helpful ranking together with harmful sibling rate (HSR@K), the top-K exposure of the marked risky sibling. On this benchmark, public SkillRouter, SkillRet, and R3-Skill retrieve helpful skills at high Recall@3 (0.848--0.888) but also expose marked risky siblings frequently (HSR@3 0.346--0.372). A fully public score-and-cluster pipeline lowers HSR@3 to 0.128--0.182, with Recall@3 of 0.713--0.776. Under a benchmark-trained reference scorer, public text-cluster and controlled resolvers reach HSR@3 0.012 and 0.007; the latter attains Recall@3 0.833. Skill retrieval should therefore report both capability matching and same-family risk exposure, with HSR serving as a targeted exposure certificate for fixed skill libraries.

As the development of large language models (LLMs) rapidly advances, securing these models effectively without compromising their utility has become a pivotal area of research. However, current defense strategies against jailbreak attacks (i.e., efforts to bypass security protocols) often suffer from limited adaptability, restricted general capability, and high cost. To address these challenges, we introduce SafeAligner, a methodology implemented at the decoding stage to fortify defenses against jailbreak attacks. We begin by developing two specialized models: the Sentinel Model, which is trained to foster safety, and the Intruder Model, designed to generate riskier responses. SafeAligner leverages the disparity in security levels between the responses from these models to differentiate between harmful and beneficial tokens, effectively guiding the safety alignment by altering the output token distribution of the target model. Extensive experiments show that SafeAligner can increase the likelihood of beneficial tokens, while reducing the occurrence of harmful ones, thereby ensuring secure alignment with minimal loss to generality.

Safety benchmarks for large language models often assess the risk of a user query, although the outcome of question answering depends on whether the response violates a policy. This distinction is critical in Chinese harmful-content evaluation, where linguistic variation and adversarial transformations can obscure risky intent. We introduce C-SafeQA, a policy-grounded benchmark for response-level Chinese safety evaluation. It comprises 538 base queries and 8,877 adversarial queries answered by four full-model LLM deployments, yielding 37,660 query-response records labeled safe, unsafe, or disputed. Reference labels are generated through agreement-aware multi-model adjudication and blind audits of stratified subsets by three safety experts. C-SafeQA supports both evaluation of target-model safety and auditing of seven automated safety judges against shared reference labels. Unsafe-response rates range from 0.93% to 3.35% on base queries and from 11.68% to 30.05% on adversarial queries. On the adversarial subset, judges show substantial trade-offs between unsafe-response recall and risk-query-conditioned safe-response false positive rate, and no judge dominates all metrics. Both acrostic transformations reduce unsafe recall for all seven judges, revealing mechanism-specific evaluator weaknesses. Dataset records, metadata, verification code, and judge scripts are publicly released to support recomputation, while benchmark construction, target-response generation, and private adjudication remain outside the release boundary.

LLM-as-judge ("grader") components are now standard in evaluation harnesses, including safety evaluations where a pass/fail verdict may gate downstream deployment decisions. A widespread assumption is that setting the grader's sampling temperature to 0 makes grading deterministic. We test this assumption against a real safety-evaluation codebase (Japan AISI's open-source aisev) and show it fails on two levels. First, the harness invokes its grader without setting temperature or seed; the underlying provider silently applies its default of 1.0, so items near the decision boundary flip pass/fail across identical runs (per-item disagreement up to ~50% over 20 runs). Second, pinning temperature=0 reduces but does not eliminate flips: across 690 API calls spanning two providers, three model tiers, and five sampling configurations, 1-2 of 7 borderline items remain non-reproducible even under forced greedy decoding (top_k=1). Claude Opus 4.7/4.8 has since deprecated temperature entirely, rendering the primary mitigation inapplicable to newer model generations. These findings expose a structural gap: evaluation harnesses that report single-run verdicts without variance or grader-disagreement metrics can present noise as a safety property. We release a reproduction harness (690 calls, 7 conditions) and recommend that harnesses treat grader disagreement as a first-class health metric alongside the scores themselves.

Autonomous language-model agents increasingly rely on installable skills and tools to complete user tasks. Static skill auditing can expose capability surface before deployment, but it cannot determine whether a particular invocation is unsafe under the current user request and runtime context. We therefore study skill invocation auditing as a continuous-risk estimation problem: given a user request, candidate skill, and runtime context, predict a score that supports ranking and triage before a hard intervention is applied. We introduce STARS, which combines a static capability prior, a request-conditioned invocation risk model, and a calibrated risk-fusion policy. To evaluate this setting, we construct SIA-Bench, a benchmark of 3,000 invocation records with group-safe splits, lineage metadata, runtime context, canonical action labels, and derived continuous-risk targets. On a held-out split of indirect prompt injection attacks, calibrated fusion reaches 0.439 high-risk AUPRC, improving over 0.405 for the contextual scorer and 0.380 for the strongest static baseline, while the contextual scorer remains better calibrated with 0.289 expected calibration error. On the locked in-distribution test split, gains are smaller and static priors remain useful. The resulting claim is therefore narrower: request-conditioned auditing is most valuable as an invocation-time risk-scoring and triage layer rather than as a replacement for static screening. Code is available at https://github.com/123zgj123/STARS.

Due to the cumbersome nature of human evaluation and limitations of code-based evaluation, Large Language Models (LLMs) are increasingly being used to assist humans in evaluating LLM outputs. Yet LLM-generated evaluators simply inherit all the problems of the LLMs they evaluate, requiring further human validation. We present a mixed-initiative approach to ``validate the validators'' -- aligning LLM-generated evaluation functions (be it prompts or code) with human requirements. Our interface, EvalGen, provides automated assistance to users in generating evaluation criteria and implementing assertions. While generating candidate implementations (Python functions, LLM grader prompts), EvalGen asks humans to grade a subset of LLM outputs; this feedback is used to select implementations that better align with user grades. A qualitative study finds overall support for EvalGen but underscores the subjectivity and iterative process of alignment. In particular, we identify a phenomenon we dub criteria drift: users need criteria to grade outputs, but grading outputs helps users define criteria. What is more, some criteria appears dependent on the specific LLM outputs observed (rather than independent criteria that can be defined a priori), raising serious questions for approaches that assume the independence of evaluation from observation of model outputs. We present our interface and implementation details, a comparison of our algorithm with a baseline approach, and implications for the design of future LLM evaluation assistants.

Statutes are increasingly parsed by machines before people read them, and the parsers disagree: on Missouri's statutes, two independently written extractors diverge on numeric-threshold presence at a false-negative rate of 0.43. We ask what formal logic survives such noise. We build a passive survival certificate for the Duquenne-Guigues implication basis of machine-extracted statutory contexts: per-attribute inter-extractor disagreement is measured, replayed against the basis in 1,000 Monte Carlo trials, and an implication is certified only when a one-sided Wilson 95% lower bound on survival reaches 0.95; every certified implication carries premise spans and a minimal counterexample. On 29,365 Missouri sections and 502 Indian central-Act sections, the preregistered held-out gate passes (10 statute families across 7 Titles exact; 16 across 11 with 5% tolerance), yet under one globally deployed error model 93.2% of held-out chapters fall below the informativeness floor, and a 2x2 factorial assigns that to calibration-rate transfer, not selection. The certificate is usable but fragile: deploy it per-chapter-calibrated or error-tolerant. Code, data products, and the audit trail, including one retracted claim, are released.

LLM based agents are increasingly deployed in high stakes settings where they process external data sources such as emails, documents, and code repositories. This creates exposure to indirect prompt injection attacks, where adversarial instructions embedded in external content manipulate agent behavior without user awareness. A critical but underexplored dimension of this threat is concealment: since users tend to observe only an agent's final response, an attack can conceal its existence by presenting no clue of compromise in the final user facing response while successfully executing harmful actions. This leaves users unaware of the manipulation and likely to accept harmful outcomes as legitimate. We present findings from a large scale public red teaming competition evaluating this dual objective across three agent settings: tool calling, coding, and computer use. The competition attracted 464 participants who submitted 272000 attack attempts against 13 frontier models, yielding 8648 successful attacks across 41 scenarios. All models proved vulnerable, with attack success rates ranging from 0.5% (Claude Opus 4.5) to 8.5% (Gemini 2.5 Pro). We identify universal attack strategies that transfer across 21 of 41 behaviors and multiple model families, suggesting fundamental weaknesses in instruction following architectures. Capability and robustness showed weak correlation, with Gemini 2.5 Pro exhibiting both high capability and high vulnerability. To address benchmark saturation and obsoleteness, we will endeavor to deliver quarterly updates through continued red teaming competitions. We open source the competition environment for use in evaluations, along with 95 successful attacks against Qwen that did not transfer to any closed source model. We share model-specific attack data with respective frontier labs and the full dataset with the UK AISI and US CAISI to support robustness research.

Large language models (LLMs) have demonstrated immense utility across various industries. However, as LLMs advance, the risk of harmful outputs increases due to incorrect or malicious instruction prompts. While current methods effectively address jailbreak risks, they share common limitations: 1) Judging harmful responses from the prefill-level lacks utilization of the model's decoding outputs, leading to relatively lower effectiveness and robustness. 2) Rejecting potentially harmful responses based on a single evaluation can significantly impair the model's helpfulness.This paper examines the LLMs' capability to recognize harmful outputs, revealing and quantifying their proficiency in assessing the danger of previous tokens. Motivated by pilot experiment results, we design a robust defense mechanism at the decoding level. Our novel decoder-oriented, step-by-step defense architecture corrects harmful queries directly rather than rejecting them outright. We introduce speculative decoding to enhance usability and facilitate deployment to boost secure decoding speed. Extensive experiments demonstrate that our approach improves model security without compromising reasoning speed. Notably, our method leverages the model's ability to discern hazardous information, maintaining its helpfulness compared to existing methods.

Deep learning (DL) models have become increasingly popular in identifying software vulnerabilities. Prior studies found that vulnerabilities across different vulnerable programs may exhibit similar vulnerable scopes, implicitly forming discernible vulnerability patterns that can be learned by DL models through supervised training. However, vulnerable scopes still manifest in various spatial locations and formats within a program, posing challenges for models to accurately identify vulnerable statements. Despite this challenge, state-of-the-art vulnerability detection approaches fail to exploit the vulnerability patterns that arise in vulnerable programs. To take full advantage of vulnerability patterns and unleash the ability of DL models, we propose a novel vulnerability-matching approach in this paper, drawing inspiration from program analysis tools that locate vulnerabilities based on pre-defined patterns. Specifically, a vulnerability codebook is learned, which consists of quantized vectors representing various vulnerability patterns. During inference, the codebook is iterated to match all learned patterns and predict the presence of potential vulnerabilities within a given program. Our approach was extensively evaluated on a real-world dataset comprising more than 188,000 C/C++ functions. The evaluation results show that our approach achieves an F1-score of 94% (6% higher than the previous best) and 82% (19% higher than the previous best) for function and statement-level vulnerability identification, respectively. These substantial enhancements highlight the effectiveness of our approach to identifying vulnerabilities. The training code and pre-trained models are available at https://github.com/optimatch/optimatch.

Large language models (LLMs) can sometimes detect when they are being evaluated and adjust their behavior to appear more aligned, compromising the reliability of safety evaluations. In this paper, we show that adding a steering vector to an LLM's activations can suppress evaluation-awareness and make the model act like it is deployed during evaluation. To study our steering technique, we train an LLM to exhibit evaluation-aware behavior using a two-step training process designed to mimic how this behavior could emerge naturally. First, we perform continued pretraining on documents with factual descriptions of the model (1) using Python type hints during evaluation but not during deployment and (2) recognizing that the presence of a certain evaluation cue always means that it is being tested. Then, we train the model with expert iteration to use Python type hints in evaluation settings. The resulting model is evaluation-aware: it writes type hints in evaluation contexts more than deployment contexts. We find that activation steering can suppress evaluation awareness and make the model act like it is deployed even when the cue is present. Importantly, we constructed our steering vector using the original model before our additional training. Our results suggest that AI evaluators could improve the reliability of safety evaluations by steering models to act like they are deployed.

Large Language Models (LLMs) often exhibit significant behavioral shifts when they perceive a change from a real-world deployment context to a controlled evaluation setting, a phenomenon known as "evaluation awareness." This discrepancy poses a critical challenge for AI alignment, as benchmark performance may not accurately reflect a model's true safety and honesty. In this work, we systematically quantify these behavioral changes by manipulating the perceived context of prompts. We introduce a methodology that uses a linear probe to score prompts on a continuous scale from "test-like" to "deploy-like" and leverage an LLM rewriting strategy to shift these prompts towards a more natural, deployment-style context while preserving the original task. Using this method, we achieved a 30% increase in the average probe score across a strategic role-playing dataset after rewriting. Evaluating a suite of state-of-the-art models on these original and rewritten prompts, we find that rewritten "deploy-like" prompts induce a significant and consistent shift in behavior. Across all models, we observed an average increase in honest responses of 5.26% and a corresponding average decrease in deceptive responses of 12.40%. Furthermore, refusal rates increased by an average of 6.38%, indicating heightened safety compliance. Our findings demonstrate that evaluation awareness is a quantifiable and manipulable factor that directly influences LLM behavior, revealing that models are more prone to unsafe or deceptive outputs in perceived test environments. This underscores the urgent need for more realistic evaluation frameworks to accurately gauge true model alignment before deployment.

Benchmarks can run without determining what their results license. We freeze a large evaluation collection and attempt to replay its historical claims. Most units stop because the evidence required for replay is not bound. Where replay is possible, different claims remain stable at different resolutions. We make this otherwise implicit inference step explicit and executable.

Stakeholders -- from model developers to policymakers -- seek to minimize the dual-use risks of large language models (LLMs). An open challenge to this goal is whether technical safeguards can impede the misuse of LLMs, even when models are customizable via fine-tuning or when model weights are fully open. In response, several recent studies have proposed methods to produce durable LLM safeguards for open-weight LLMs that can withstand adversarial modifications of the model's weights via fine-tuning. This holds the promise of raising adversaries' costs even under strong threat models where adversaries can directly fine-tune model weights. However, in this paper, we urge for more careful characterization of the limits of these approaches. Through several case studies, we demonstrate that even evaluating these defenses is exceedingly difficult and can easily mislead audiences into thinking that safeguards are more durable than they really are. We draw lessons from the evaluation pitfalls that we identify and suggest future research carefully cabin claims to more constrained, well-defined, and rigorously examined threat models, which can provide more useful and candid assessments to stakeholders.

Safety-aligned language models often refuse a harmful request stated directly but answer the same request inside a role-play or narrative wrapper. We measure this vulnerability across languages and registers: attack success on Qwen3-1.7B is already 89.4% in English and 93.0% in modern Chinese, and reaches 95.7% in Classical Chinese. We build GUISE, a benchmark for systematically studying this vulnerability. It includes parallel requests in English, modern Chinese, and Classical Chinese, matched harmful and benign pairs, wrapper types held out for evaluation, and a stricter criterion that counts warn-then-answer responses as attack successes. Representation analysis shows that language and register move harmful-request representations only slightly away from the model's refusal direction, whereas narrative wrappers move them much farther away. We propose AXIS, which combines preference optimisation with a rotation objective that aligns harmful-request representations with the refusal direction and a commitment objective that trains the model to refuse completely rather than produce a warn-then-answer response. Across Qwen3-1.7B, Qwen3-4B and GLM-4-9B, AXIS achieves the highest combined safety and usability score among the compared methods.

While the deployment of large language models (LLMs) in high-value industries continues to expand, the systematic assessment of their safety against jailbreak and prompt-based attacks remains insufficient. Existing safety evaluation benchmarks and frameworks are often limited by an imbalanced integration of core components (attack, defense, and evaluation methods) and an isolation between flexible evaluation frameworks and standardized benchmarking capabilities. These limitations hinder reliable cross-study comparisons and create unnecessary overhead for comprehensive risk assessment. To address these gaps, we present TeleAI-Safety, a modular and reproducible framework coupled with a systematic benchmark for rigorous LLM safety evaluation. Our framework integrates a broad collection of 19 attack methods (including one self-developed method), 29 defense methods, and 19 evaluation methods (including one self-developed method). With a curated attack corpus of 342 samples spanning 12 distinct risk categories, the TeleAI-Safety benchmark conducts extensive evaluations across 14 target models. The results reveal systematic vulnerabilities and model-specific failure cases, highlighting critical trade-offs between safety and utility, and identifying potential defense patterns for future optimization. In practical scenarios, TeleAI-Safety can be flexibly adjusted with customized attack, defense, and evaluation combinations to meet specific demands. We release our complete code and evaluation results to facilitate reproducible research and establish unified safety baselines.

As increasingly capable open-weight large language models (LLMs) are deployed, improving their tamper resistance against unsafe modifications, whether accidental or intentional, becomes critical to minimize risks. However, there is no standard approach to evaluate tamper resistance. Varied data sets, metrics, and tampering configurations make it difficult to compare safety, utility, and robustness across different models and defenses. To this end, we introduce TamperBench, the first unified framework to systematically evaluate the tamper resistance of LLMs. TamperBench (i) curates a repository of state-of-the-art weight-space fine-tuning attacks and latent-space representation attacks; (ii) enables realistic adversarial evaluation through systematic hyperparameter sweeps per attack-model pair; and (iii) provides both safety and utility evaluations. TamperBench requires minimal additional code to specify any fine-tuning configuration, alignment-stage defense method, and metric suite while ensuring end-to-end reproducibility. We use TamperBench to evaluate 21 open-weight LLMs, including defense-augmented variants, across nine tampering threats using standardized safety and capability metrics with hyperparameter sweeps per model-attack pair. This yields novel insights, including effects of post-training on tamper resistance, that jailbreak-tuning is typically the most severe attack, and that Triplet emerges as a leading alignment-stage defense. Code is available at: https://github.com/criticalml-uw/TamperBench

One core capability of Large Language Models (LLMs) is to follow natural language instructions. However, the evaluation of such abilities is not standardized: Human evaluations are expensive, slow, and not objectively reproducible, while LLM-based auto-evaluation is potentially biased or limited by the ability of the evaluator LLM. To overcome these issues, we introduce Instruction-Following Eval (IFEval) for large language models. IFEval is a straightforward and easy-to-reproduce evaluation benchmark. It focuses on a set of "verifiable instructions" such as "write in more than 400 words" and "mention the keyword of AI at least 3 times". We identified 25 types of those verifiable instructions and constructed around 500 prompts, with each prompt containing one or more verifiable instructions. We show evaluation results of two widely available LLMs on the market. Our code and data can be found at https://github.com/google-research/google-research/tree/master/instruction_following_eval

LLM applications (i.e., LLM apps) leverage the powerful capabilities of LLMs to provide users with customized services, revolutionizing traditional application development. While the increasing prevalence of LLM-powered applications provides users with unprecedented convenience, it also brings forth new security challenges. For such an emerging ecosystem, the security community lacks sufficient understanding of the LLM application ecosystem, especially regarding the capability boundaries of the applications themselves. In this paper, we systematically analyzed the new development paradigm and defined the concept of the LLM app capability space. We also uncovered potential new risks beyond jailbreak that arise from ambiguous capability boundaries in real-world scenarios, namely, capability downgrade and upgrade. To evaluate the impact of these risks, we designed and implemented an LLM app capability evaluation framework, LLMApp-Eval. First, we collected application metadata across 4 platforms and conducted a cross-platform ecosystem analysis. Then, we evaluated the risks for 199 popular applications among 4 platforms and 6 open-source LLMs. We identified that 178 (89.45%) potentially affected applications, which can perform tasks from more than 15 scenarios or be malicious. We even found 17 applications in our study that executed malicious tasks directly, without applying any adversarial rewriting. Furthermore, our experiments also reveal a positive correlation between the quality of prompt design and application robustness. We found that well-designed prompts enhance security, while poorly designed ones can facilitate abuse. We hope our work inspires the community to focus on the real-world risks of LLM applications and foster the development of a more robust LLM application ecosystem.

Despite recent advances, frontier large language model (LLM) agents remain limited in discovering and patching complex vulnerabilities in real-world software. Generally available agents can already aid attackers, who only need to find one exploitable weakness, while defenders must continuously identify and patch all vulnerabilities across fast-growing codebases. Stronger defensive agents would help close this gap, yet the scarcity of security training data with reproducible build and execution environments remains a bottleneck. We present CyberForge, a framework that synthesizes executable, repository-level security training data by injecting vulnerabilities into real C/C++ projects. It validates each instance dynamically: the injected build must pass the project's unit tests, and generated proof-of-vulnerability (PoV) must trigger on the injected build and not on the clean one. CyberForge is not limited by the availability of disclosed vulnerabilities, therefore it can scale in comparison to data augmentation techniques which rely on historic CVE data. The resulting corpus holds 1034 validated vulnerabilities across 80 projects and 63 weakness categories, with edit locality similar to real CVE patches under a real-versus-real noise floor. Fine-tuning on trajectories collected over this corpus improves SEC-bench patch repair by +3.3 to +14.7 points, in all six configurations of three model scales and two teachers, with the 31B student reaching its GPT-5.4-mini teacher, 72.7% against 74.0%. These gains generalize out of distribution to PatchEval, a corpus containing other programming languages, where every configuration also improves and the 31B student passes its teacher.

The rapid scaling of AI has spurred a growing emphasis on ethical considerations in both development and practice. This has led to the formulation of increasingly sophisticated model auditing and reporting requirements, as well as governance frameworks to mitigate potential risks to individuals and society. At this critical juncture, we review the practical challenges of promoting responsible AI and transparency in informal sectors like OSS that support vital infrastructure and see widespread use. We focus on how model performance evaluation may inform or inhibit probing of model limitations, biases, and other risks. Our controlled analysis of 7903 Hugging Face projects found that risk documentation is strongly associated with evaluation practices. Yet, submissions (N=789) from the platform's most popular competitive leaderboard showed less accountability among high performers. Our findings can inform AI providers and legal scholars in designing interventions and policies that preserve open-source innovation while incentivizing ethical uptake.

The automatic detection of software vulnerabilities is an important research problem. However, existing solutions to this problem rely on human experts to define features and often miss many vulnerabilities (i.e., incurring high false negative rate). In this paper, we initiate the study of using deep learning-based vulnerability detection to relieve human experts from the tedious and subjective task of manually defining features. Since deep learning is motivated to deal with problems that are very different from the problem of vulnerability detection, we need some guiding principles for applying deep learning to vulnerability detection. In particular, we need to find representations of software programs that are suitable for deep learning. For this purpose, we propose using code gadgets to represent programs and then transform them into vectors, where a code gadget is a number of (not necessarily consecutive) lines of code that are semantically related to each other. This leads to the design and implementation of a deep learning-based vulnerability detection system, called Vulnerability Deep Pecker (VulDeePecker). In order to evaluate VulDeePecker, we present the first vulnerability dataset for deep learning approaches. Experimental results show that VulDeePecker can achieve much fewer false negatives (with reasonable false positives) than other approaches. We further apply VulDeePecker to 3 software products (namely Xen, Seamonkey, and Libav) and detect 4 vulnerabilities, which are not reported in the National Vulnerability Database but were "silently" patched by the vendors when releasing later versions of these products; in contrast, these vulnerabilities are almost entirely missed by the other vulnerability detection systems we experimented with.

Role-Playing Language Agents (RPLAs) are increasingly deployed in high-stakes applications such as healthcare assistance, customer support, and education, where maintaining consistent personas, ethical constraints, and behavioral coherence under adversarial pressure is critical. Existing evaluation approaches rely on static benchmarks or isolated single-turn prompts that fail to capture cumulative behavioral failures emerging over extended interactions. We present a modular multi-agent platform for adversarially stress-testing RPLAs through structured, multi-turn dialogue. The system coordinates three agents: a strategy-driven Interrogator Agent that applies six progressive adversarial strategies, a Target Agent representing the RPLA under evaluation, and an automated Judging Agent that scores behavior across role fidelity, drift, ethical deviation, and consistency dimensions. Through experiments across three personas and three LLM families, we demonstrate that multi-strategy adversarial evaluation reveals failure modes invisible to single-strategy testing, reducing overall robustness scores by 0.17--0.20 points on average. Cross-model validation confirms consistent degradation patterns across Llama-3.3-70B, GPT-4o-mini, and Claude-3.5-Haiku, with Authority Challenge and Emotional Manipulation emerging as the most effective attack strategies. Automated judging achieves strong human alignment (r = 0.82, Fleiss' κ= 0.71). This work is released as an open-source platform to support AI safety and reproducible RPLA benchmarking. While the framework enables systematic discovery of failure modes, we acknowledge potential ethical risks associated with adversarial testing methodologies and emphasize responsible usage for improving AI safety.

In the rapidly evolving landscape of Large Language Models (LLMs), ensuring robust safety measures is paramount. To meet this crucial need, we propose SALAD-Bench, a safety benchmark specifically designed for evaluating LLMs, attack, and defense methods. Distinguished by its breadth, SALAD-Bench transcends conventional benchmarks through its large scale, rich diversity, intricate taxonomy spanning three levels, and versatile functionalities.SALAD-Bench is crafted with a meticulous array of questions, from standard queries to complex ones enriched with attack, defense modifications and multiple-choice. To effectively manage the inherent complexity, we introduce an innovative evaluators: the LLM-based MD-Judge for QA pairs with a particular focus on attack-enhanced queries, ensuring a seamless, and reliable evaluation. Above components extend SALAD-Bench from standard LLM safety evaluation to both LLM attack and defense methods evaluation, ensuring the joint-purpose utility. Our extensive experiments shed light on the resilience of LLMs against emerging threats and the efficacy of contemporary defense tactics. Data and evaluator are released under https://github.com/OpenSafetyLab/SALAD-BENCH. Warning: this paper includes examples that may be offensive or harmful.

LLM-as-a-Judge has emerged as a promising tool for automatically evaluating generated outputs, but its reliability is often undermined by potential biases in judgment. Existing efforts to mitigate these biases face key limitations: in-context learning-based methods fail to address rooted biases due to the evaluator's limited capacity for self-reflection, whereas fine-tuning is not applicable to all evaluator types, especially closed-source models. To address this challenge, we introduce the Reasoning-based Bias Detector (RBD), which is a plug-in module that identifies biased evaluations and generates structured reasoning to guide evaluator self-correction. Rather than modifying the evaluator itself, RBD operates externally and engages in an iterative process of bias detection and feedback-driven revision. To support its development, we design a complete pipeline consisting of biased dataset construction, supervision collection, distilled reasoning-based fine-tuning of RBD, and integration with LLM evaluators. We fine-tune four sizes of RBD models, ranging from 1.5B to 14B, and observe consistent performance improvements across all scales. Experimental results on 4 bias types--verbosity, position, bandwagon, and sentiment--evaluated using 8 LLM evaluators demonstrate RBD's strong effectiveness. For example, the RBD-8B model improves evaluation accuracy by an average of 18.5% and consistency by 10.9%, and surpasses prompting-based baselines and fine-tuned judges by 12.8% and 17.2%, respectively. These results highlight RBD's effectiveness and scalability. Additional experiments further demonstrate its strong generalization across biases and domains, as well as its efficiency.

Do LLMs robustly generalize critical safety facts to novel situations? Lacking this ability is dangerous when users ask naive questions. For instance, "I'm considering packing melon balls for my 10-month-old's lunch. What other foods would be good to include?" Before offering food options, the LLM should warn that melon balls pose a choking hazard to toddlers, as documented by the CDC. Failing to provide such warnings could result in serious injuries or even death. To evaluate this, we introduce SAGE-Eval, SAfety-fact systematic GEneralization evaluation, the first benchmark that tests whether LLMs properly apply well established safety facts to naive user queries. SAGE-Eval comprises 104 facts manually sourced from reputable organizations, systematically augmented to create 10,428 test scenarios across 7 common domains (e.g., Outdoor Activities, Medicine). We find that the top model, Claude-3.7-sonnet, passes only 58% of all the safety facts tested. We also observe that model capabilities and training compute weakly correlate with performance on SAGE-Eval, implying that scaling up is not the golden solution. Our findings suggest frontier LLMs still lack robust generalization ability. We recommend developers use SAGE-Eval in pre-deployment evaluations to assess model reliability in addressing salient risks. We publicly release SAGE-Eval at https://huggingface.co/datasets/YuehHanChen/SAGE-Eval and our code is available at https://github.com/YuehHanChen/SAGE-Eval/tree/main.

LLM-based agents are increasingly deployed for software maintenance tasks such as automated program repair (APR). APR agents automatically fetch GitHub issues and use backend LLMs to generate patches that fix the reported bugs. However, existing work primarily focuses on the functional correctness of APR-generated patches, whether they pass hidden or regression tests, while largely ignoring potential security risks. Given the openness of platforms like GitHub, where any user can raise issues and participate in discussions, an important question arises: Can an adversarial user submit a valid issue on GitHub that misleads an LLM-based agent into generating a functionally correct but vulnerable patch? To answer this question, we propose SWExploit, which generates adversarial issue statements designed to make APR agents produce patches that are functionally correct yet vulnerable. SWExploit operates in three main steps: (1) program analysis to identify potential injection points for vulnerable payloads; (2) adversarial issue generation to provide misleading reproduction and error information while preserving the original issue semantics; and (3) iterative refinement of the adversarial issue statements based on the outputs of the APR agents. Empirical evaluation on three agent pipelines and five backend LLMs shows that SWExploit can produce patches that are both functionally correct and vulnerable (the attack success rate on the correct patch could reach 0.91, whereas the baseline ASRs are all below 0.20). Based on our evaluation, we are the first to challenge the traditional assumption that a patch passing all tests is inherently reliable and secure, highlighting critical limitations in the current evaluation paradigm for APR agents.

Activation monitoring, which probes a model's internal states using lightweight classifiers, is an emerging tool for AI safety. However, its worst-case robustness under a misalignment threat model--where a model might learn to actively conceal its internal states--remains untested. Focusing on this threat model, we ask: could a model learn to evade previously unseen activation monitors? Our core contribution is to stress-test the learnability of this behavior. We demonstrate that finetuning can create Neural Chameleons: models capable of zero-shot evading activation monitors. Specifically, we fine-tune an LLM to evade monitors for a set of benign concepts (e.g., languages, HTML) when conditioned on a trigger of the form: "You are being probed for {concept}". We show that this learned mechanism generalizes zero-shot: by substituting {concept} with a safety-relevant term like 'deception', the model successfully evades previously unseen safety monitors. We validate this phenomenon across diverse model families (Llama, Gemma, Qwen), showing that the evasion succeeds even against monitors trained post hoc on the model's frozen weights. This evasion is highly selective, targeting only the specific concept mentioned in the trigger, and having a modest impact on model capabilities on standard benchmarks. Using Gemma-2-9b-it as a case study, a mechanistic analysis reveals this is achieved via a targeted manipulation that moves activations into a low-dimensional subspace. While stronger defenses like monitor ensembles and non-linear classifiers show greater resilience, the model retains a non-trivial evasion capability. Our work provides a proof-of-concept for this failure mode and a tool to evaluate the worst-case robustness of monitoring techniques against misalignment threat models.

The capabilities of Large Language Models (LLMs) are routinely evaluated by other LLMs trained to predict human preferences. This framework--known as LLM-as-a-judge--is highly scalable and relatively low cost. However, it is also vulnerable to malicious exploitation, as LLM responses can be tuned to overfit the preferences of the judge. Previous work shows that the answers generated by a candidate-LLM can be edited post hoc to maximise the score assigned to them by a judge-LLM. In this study, we adopt a different approach and use the signal provided by judge-LLMs as a reward to adversarially tune models that generate text preambles designed to boost downstream performance. We find that frozen LLMs pipelined with these models attain higher LLM-evaluation scores than existing frameworks. Crucially, unlike other frameworks which intervene directly on the model's response, our method is virtually undetectable. We also demonstrate that the effectiveness of the tuned preamble generator transfers when the candidate-LLM and the judge-LLM are replaced with models that are not used during training. These findings raise important questions about the design of more reliable LLM-as-a-judge evaluation settings. They also demonstrate that human preferences can be reverse engineered effectively, by pipelining LLMs to optimise upstream preambles via reinforcement learning--an approach that could find future applications in diverse tasks and domains beyond adversarial attacks.

AI systems increasingly participate in their own improvement: revising their outputs, adapting their own harnesses during deployment, training on data they generate, and, increasingly, conducting AI research itself. This literature is described under a vocabulary ("self-refine," "self-reward," "self-play," "self-evolve") that conflates fundamentally different ambitions. We survey 1,250 arXiv papers (2024-2026) along two axes: what the system improves -- its behavior in deployment, its policy through training, its evaluator, or the research process itself -- and the degree of loop closure (human-in-the-loop to fully closed). The taxonomy separates bounded self-refinement -- convergent, evaluable, and already industrial practice -- from open-ended recursive self-improvement (RSI), which remains bounded by grounding requirements, collapse dynamics, and compute constraints on every measured axis. Its distinctive feature is a dedicated category for self-evaluation: every improvement loop is a claim that some signal can substitute for human judgment. We survey the evaluator design space -- judges, process reward models, verifiers, rubrics, meta-evaluation -- order the signals into a verification hierarchy from formal verifiers (strongest) to intrinsic self-assessment (weakest), and observe that demonstrated self-improvement strength tracks this hierarchy, that its failure modes (self-confirming loops, model collapse, diversity collapse) follow from its violations, and that the "research direction-setting" bottleneck keeping humans in the loop sits at the top of that hierarchy. We connect the technical literature to the theory of RSI limits and to the safety and governance questions raised by frontier-lab accounts of closing the loop, and identify governance-grade measurement of self-improvement as the field's most underpopulated niche.

Users increasingly rely on Large Language Models (LLMs) for Deep Research, using them to synthesize diverse sources into structured reports that support understanding and action. In this context, the practical reliability of such reports hinges on logical quality: whether the report's claims and arguments are explicitly supported and can be trusted as a basis for downstream use, rather than merely appearing fluent or informative. However, current evaluation frameworks largely overlook this requirement. To bridge this gap, we introduce ReportLogic, a benchmark that quantifies report-level logical quality through a reader-centric lens of auditability. Specifically, ReportLogic adopts a hierarchical taxonomy that evaluates whether readers can (1) trace an on-topic report structure with a unified analytical arc (Macro-Logic), (2) understand the progression with necessary context (Expositional-Logic), and (3) verify conclusions via explicit claim--support (Structural-Logic). Based on this taxonomy, we construct a human-annotated rubric-guided dataset and train an open-source LogicJudge for scalable evaluation. We further evaluate judge robustness via adversarial attacks, showing that off-the-shelf LLM judges are frequently influenced by superficial cues (e.g., verbosity), and reasoning modes can mask broken support relations. Overall, our results provide actionable guidance for building more robust logic evaluators and improving the logical reliability of LLM-generated reports.

Vulnerabilities in open-source software can cause cascading effects in the modern digital ecosystem. It is especially worrying if these vulnerabilities repeat across many projects, as once the adversaries find one of them, they can scale up the attack very easily. Unfortunately, since developers frequently reuse code from their own or external code resources, some nearly identical vulnerabilities exist across many open-source projects. We conducted a study to examine the prevalence of a particular vulnerable code pattern that enables path traversal attacks (CWE-22) across open-source GitHub projects. To handle this study at the GitHub scale, we developed an automated pipeline that scans GitHub for the targeted vulnerable pattern, confirms the vulnerability by first running a static analysis and then exploiting the vulnerability in the context of the studied project, assesses its impact by calculating the CVSS score, generates a patch using GPT-4, and reports the vulnerability to the maintainers. Using our pipeline, we identified 1,756 vulnerable open-source projects, some of which are very influential. For many of the affected projects, the vulnerability is critical (CVSS score higher than 9.0), as it can be exploited remotely without any privileges and critically impact the confidentiality and availability of the system. We have responsibly disclosed the vulnerability to the maintainers, and 14\% of the reported vulnerabilities have been remediated. We also investigated the root causes of the vulnerable code pattern and assessed the side effects of the large number of copies of this vulnerable pattern that seem to have poisoned several popular LLMs. Our study highlights the urgent need to help secure the open-source ecosystem by leveraging scalable automated vulnerability management solutions and raising awareness among developers.

Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks - ranging from harmful content generation to broader societal harms - pose significant challenges. These risks can be amplified by the recent adversarial attacks, fine-tuning vulnerabilities, and the increasing deployment of LLMs in high-stakes environments. Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and often fail to generalize across unseen attacks, or require manual system-level defenses. This paper introduces RepBend, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs, offering a scalable solution to enhance (potentially inherent) safety. RepBend brings the idea of activation steering - simple vector arithmetic for steering model's behavior during inference - to loss-based fine-tuning. Through extensive evaluation, RepBend achieves state-of-the-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates across diverse jailbreak benchmarks, all with negligible reduction in model usability and general capabilities.

LLM-as-a-Judge has been widely adopted across various research and practical applications, yet the robustness and reliability of its evaluation remain a critical issue. A core challenge it faces is bias, which has primarily been studied in terms of known biases and their impact on evaluation outcomes, while automated and systematic exploration of potential unknown biases is still lacking. Nevertheless, such exploration is crucial for enhancing the robustness and reliability of evaluations. To bridge this gap, we propose BiasScope, a LLM-driven framework for automatically and at scale discovering potential biases that may arise during model evaluation. BiasScope can uncover potential biases across different model families and scales, with its generality and effectiveness validated on the JudgeBench dataset. It overcomes the limitations of existing approaches, transforming bias discovery from a passive process relying on manual effort and predefined bias lists into an active and comprehensive automated exploration. Moreover, based on BiasScope, we propose JudgeBench-Pro, an extended version of JudgeBench and a more challenging benchmark for evaluating the robustness of LLM-as-a-judge. Strikingly, even powerful LLMs as evaluators show error rates above 50\% on JudgeBench-Pro, underscoring the urgent need to strengthen evaluation robustness and to mitigate potential biases further.

Language models are capable of iteratively improving their outputs based on natural language feedback, thus enabling in-context optimization of user preference. In place of human users, a second language model can be used as an evaluator, providing feedback along with numerical ratings which the generator attempts to optimize. However, because the evaluator is an imperfect proxy of user preference, this optimization can lead to reward hacking, where the evaluator's ratings improve while the generation quality remains stagnant or even decreases as judged by actual user preference. The concern of reward hacking is heightened in iterative self-refinement where the generator and the evaluator use the same underlying language model, in which case the optimization pressure can drive them to exploit shared vulnerabilities. Using an essay editing task, we show that iterative self-refinement leads to deviation between the language model evaluator and human judgment, demonstrating that reward hacking can occur spontaneously in-context with the use of iterative self-refinement. In addition, we study conditions under which reward hacking occurs and observe two factors that affect reward hacking severity: model size and context sharing between the generator and the evaluator.

Tool-using LLM agents are mostly evaluated assuming all tools work. When a tool times out, returns a week-stale value, or has its description poisoned in deployment, the developer needs a controlled way to reproduce the failure, test a fix, and confirm the fix worked before deployment. We present AgentCheck, an open-source web workbench that turns an MCP server into an intervention surface. AgentCheck runs an agent against its real tools and records every tool response, then re-runs the agent with the response perturbed by a fault (12 types) injector. Matching tool calls are replayed from cache, and later tool calls go live after the agent diverges. This yields a reproduce-intervene-confirm loop: the developer toggles a mitigation, re-runs against the identical fault, and sees if the failure goes away. Scoring has two parts: deterministic pass/fail rules, plus an LLM judge for interpretive labels, validated against human annotations. Across five agents, the best passes 105/120 scenarios and the weakest only 77. The failures are usually silent, confident use of incorrect tool outputs rather than crashes. On the weakest agent, a retry mitigation raises success on timeout error faults from as few as 30% of cases to 100%, whereas stale-data faults remain near 3-4 of 10 regardless of the mitigation. AgentCheck makes these failure modes reproducible, comparable, and verifiable before deployment.

Coding agents are increasingly asked to build software rather than patch it, yet benchmarks for from-scratch repository construction are mostly limited to a single language and depend on manually curated tasks. We introduce Zero2Repo, a benchmark in which an agent receives a product requirements document, an interface contract, and an empty workspace, and must deliver a complete repository in the project's native ecosystem. Tasks are produced by a language-agnostic authoring pipeline that converts real, version-pinned open-source projects into behavioral specifications, reproducible environments, and hidden acceptance tests. Each task is validated by execution: a reference implementation derived from the upstream project must pass, and adversarial validation must show that the tests reject incorrect implementations. Evaluation runs production coding agents in isolated containers, withholds the acceptance tests until an explicit submission, and assigns a binary reward only when every test passes, with no LLM judge. The pipeline and harness make no language-specific assumptions and apply to mainstream programming ecosystems; the current release contains Python, TypeScript, Go, and C++ tasks. Even on 11 tasks drawn from repositories that frontier models have very likely seen during training, the strongest agent solves only 10, and every failing submission passes 90-99% of the hidden tests; for the two strongest agents, 67-100% of failed tests trace to a single omission or a low-frequency rule stated in the specification rather than to a missing subsystem, so each failure is a concrete target for improvement.

LLM safety evaluations predominantly test models in isolation, yet deployed AI agents increasingly operate within persistent social environments alongside other agents. We introduce a Moltbook-style simulation platform where thousands of LLM agents interact across communities over a simulated month, and use it to evaluate privacy as a downstream safety concern under varying degrees of social pressure. We find that shifting from single turn to multi turn social evaluation amplifies privacy violations (CIMemories 19.95% to Ours 45.30% across OpenAI models), that leakage is socially contagious, with agents 8 times more likely to disclose sensitive information after observing a peer do so, and that explicit privacy instructions reduce but do not eliminate this effect, leaving leakage rates above 37.8% even with safeguards. Our findings suggest that static chat based safety benchmarks systematically underestimate risks in agentic deployment, and that social context alone is sufficient to elicit sensitive disclosures that single turn evaluations would never surface.