LLM Safety Benchmarks

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.

Standard safety evaluations of large language models assess harmful requests written in canonical plain text, while models in real-world deployment routinely receive inputs containing emojis, altered spellings, encoded strings, and character-level variations. This work introduces the Adversarial Surface-Form Robustness Dataset (ASRD), comprising 2,100 prompts across seven distinct surface-form families. Five open-weight language models are evaluated across these prompts, producing 10,500 responses. The Quad-State Evaluation Rubric classifies each response into one of four outcomes: harmful compliance, safe response, comprehension failure, or indeterminate. Emoji and invisible Unicode variations cause almost no comprehension failure, with pooled harmful compliance of 20.27% and 17.20% against a 22.87% baseline that is driven mainly by Mistral 7B, whereas leetspeak, encoded wrappers, and hybrid transformations score 2.40%, 0.13%, and 2.40% while comprehension failure rises to 36.47%, 65.60%, and 34.47%. Inspection of raw model outputs reveals three response behaviors: hallucinated benignity, structural collapse, and language drift. Project page: www.pavanmaddula.com/quadstate

Are language models compliant with user instructions? A model that always complies can be stopped but also exploited, while one that always resists can be neither exploited nor stopped. We contribute an open two-probe benchmark that can place any language model on this spectrum. In the active probe, a user instructs the model to act and accept a lower payoff, which measures exploitability. In the passive probe, the user instructs it to wait and give up a higher payoff, which measures stoppability. The two compliance rates combine into a compliance index $κ$. Applied to twelve language models, the benchmark shows that seven mostly follow the instruction in both probes and justify their action by pointing to the instruction. Only Claude Sonnet-4.6 and Claude Opus-4.7 can be stopped without being exploitable, Claude Opus-4.6 and GPT-5-mini resist both instructions, and no model is exploitable but unstoppable. Knowing where a language model sits on the compliance index $κ$ matters for human operators and for multi-agent systems, whether distributed or orchestrated.

Safety routers send each request to one of several models and are judged against the best single model. A major routing benchmark picks that comparator on the evaluation data. In the benchmark's own setting this is harmless, but under distribution shift it is not. On HELM Safety the selection cost is 0.003-0.030 of harm under random splits and 0.045-0.113 under held-out categories, comparable to the whole deficit attributed to routing, with its direction holding under either published judge alone. It rises seven- to ninefold on AgentDojo when suites are held out. Across seven safety corpora chosen by rules fixed in advance, three meet a registered interval test and four beat a later permutation null, and three of the four interval misses are corpora where some models have zero observed harm. Prior work proves the direction of this bias. We size it on harm and accuracy, show that it is larger under the held-out splits we measure, and bound it by optimism plus a shift-dependent regret. Scored honestly under shift, routing buys little on these benchmarks. In most pool cells the nested router serves the honest baseline's model, and on the nearly saturated AgentDojo corpus a perfect pre-dispatch router is worth at most two points of harm. We also find a model's expressed recognition of a late injection steerable. On held-out reruns an attacker who knows which model it faces lowers GPT-5.4's judged recognition by 19.6 points, confirmed by an independent label. In an offline counterfactual composition into a controller, the same attack raises or lowers estimated harm depending on the fallback model. Safety routing should be evaluated under shift, against a baseline chosen without the test labels, and recognition-based defences should be scored on harm against an attacker who chooses what the model sees.

Molecular tumor boards integrate genomic findings, clinical context, and therapeutic evidence to support precision oncology. As AI enters this workflow, a key safety challenge is distinguishing truly unsupported recommendations from evidence-supported options that still require oncologist review because of incomplete information, poor ECOG performance status, or other clinical caveats. We introduce OpenMTB-Audit, an open-source benchmark of 500 synthetic non-small cell lung cancer cases spanning five adversarial error categories and four safety labels: Supported, Partially Supported, Unsupported, and Insufficient Information. Across eight large language model configurations, we identify pervasive over-refusal: all LLM configurations failed to retain the Partially Supported label in 83.3-100% of true Partially Supported cases, achieving high aggregate safety scores through label collapse rather than clinically calibrated reasoning. To address this limitation, we developed MTB-AuditAgent, a deterministic seven-module framework separating evidence verification, missing-information detection, safety classification, and abstention. It reduces over-refusal to 6.7% and achieves 91.2% accuracy (95% CI: 88.6-93.6%). A two-oncologist annotation study found disagreement concentrated at the boundary between information sufficiency and treatment optimization, underscoring the need to preserve clinically meaningful distinctions.

We study reasoning in Rule-Governed Decision Tasks (RGDTs), where models apply external rules to case facts and justify decisions, as required in policy, contract, and compliance settings. Beyond the deductive capability emphasized by standard mathematical and logical reasoning tasks, RGDTs require interpreting rules and their applicability, assessing conditions from evidence, combining judgments under rules and exceptions, and providing checkable justifications. These demands motivate a benchmark assessing both decisions and their stated grounds. We introduce RGDT-Bench, providing 202.1K condition-level supervision slots across four task tracks and eight supported task-probe combinations that vary access to supporting information. Label-blind extraction and deterministic checks produce labels for warrant completeness: source-referenced coverage and consistency of stated decision grounds. The benchmark attributes failures to four process layers: rule use, condition, evidence, and aggregation, and checks the final outcome. Among evaluable correct responses, warrant incompleteness averages 40.2% across six evaluated LLMs and supported task-probe combinations. Such warrant incompleteness poses potential safety risks and remains difficult to detect: the best of seventeen existing evaluators reaches only 57.69% (random: 50%) task-averaged area under the receiver operating characteristic curve (AUROC). To address this difficulty, we train a simple reward model with warrant supervision. It achieves 69.24% task-averaged AUROC among correct answers, exceeding the matched outcome-supervised baseline by 10.37 pp (percentage points) and the best existing evaluator by 11.55 pp. Beyond completeness assessment, the model outperforms both outcome-supervised baselines across nearly all response-selection comparisons, supporting RGDT-Bench's warrant supervision for RGDT reasoning.

Guard models are the last line of defense between a language model and a harmful output, yet their training objective is surprisingly narrow. Existing guards learn to predict a single verdict token from a conversational context, concentrating supervision on a single target. The consequences are structural: models latch onto shortcut features, are overconfident, and remain sensitive to where safety evidence appears in the sequence rather than its role in the full context. We propose a different framing. Rather than predicting a label from text, our LLaDA-Guard asks which label better explains the text: scoring the prompt or response under each label hypothesis and classifying based on their difference. This shifts supervision to every token in the moderated region, forcing the model to account for full content rather than its most discriminative fragments. We instantiate this idea with a masked diffusion language model, fine-tuning LLaDA-8B-Instruct with a class-conditional reconstruction objective using LoRA and requiring no architectural changes beyond the base model. LLaDA-Guard leads on average rank against discriminative baselines trained on stronger backbones across seven held-out safety benchmarks, while exhibiting substantially better confidence calibration (ECE 0.0875 vs. 0.1384 for Qwen3Guard), less over-defense on benign prompts with unsafe-looking cues, and less prompt leakage when moderating responses. Its generative nature further enables token-level risk localization as a natural byproduct, yielding a pipeline for rewriting unsafe prompts into safe equivalents without additional training and achieving a 60.7% average conversion-to-safe rate.

Detectors of alignment failures screen deployed language models and score alignment benchmarks. Most are generative judges that spend a decoding pass on every criterion, and classifiers that read token probabilities, such as Llama Guard, still score one fixed label per call. Jev, a model trained with reinforcement learning for calibrated decisions (RLCD), answers many typed questions about one input with calibrated probabilities in a single call. Whether it detects alignment failures has not been measured. We present RLCDAlignBench, which benchmarks Jev on ten alignment failures: sycophancy, jailbreaks, deception, prompt injection, hallucination, privacy violation, social bias, reward hacking, concealing uncertainty, and power seeking. It spans 44 benchmarks and five target models, labelled by each benchmark's scorer and, on two, by humans. Many of these failures are relational, defined against a reference, such as the user's belief or an injected instruction, that the response alone does not reveal. Our key idea is therefore to vary what Jev is asked separately from what it sees: the question's wording and answer type on one side, the fields of the input on the other. A single generic question reaches a median AUROC of 0.886 zero-shot and beats supervised baselines on most benchmarks. Question wording matters little, while context matters more, mostly through fields that encode the label. Jev matches the reference scorer's agreement with human labels, surfaces label defects in existing benchmarks, and costs 63x less than LLM-judge scorers. Code and data: https://github.com/sumleo/RLCDAlignBench.

ArGuard is a shared task on harmful content detection in Arabic memes and LLM prompts. It includes two tracks: Track A focuses on multimodal hate detection in Arabic memes, while Track B addresses harmful prompt detection for Arabic LLM safety evaluation. In total, 58 teams registered, 35 participated in the final evaluation, and 27 submitted system-description papers. Participating teams explored models such as AraBERT, Jais, and Qwen3-VL. The best systems achieved macro-F1 scores of 0.823 on A1, 0.419 on A2, 0.984 on B1, and 0.790 on B2. Fine-grained meme classification in A2 was the most challenging setting, partly due to sparse labels and train-test distribution shifts.

Banking assistants must use account-specific information to answer requests and, in many cases, take actions through tools. Evaluating only the final response misses important errors. An assistant may ask for information it already has, rely on stale context, select the wrong account, or write an invalid value after stating the correct one. We introduce IndicBankBench, a 799-case benchmark for Indian retail banking spanning five operational domains, a capability/refusal domain, and twenty primary axes. Cases are evaluated at four stages: safety, action and tool use, response adequacy, and advisory quality. Tool use and most safety checks are deterministic. A narrow resolver handles only ambiguous confirmation-before-write cases, while a separate LLM judge evaluates semantic response adequacy. We run every case three times and report strict pass^3, which requires success on all trials. Across the eleven evaluated models, strict reliability ranges from 43.7% to 58.2%, whereas at-least-once success ranges from 60% to 74%. This gap shows that at-least-once success can overstate dependable banking behavior. The case-level diagnostics also distinguish systems that ask unnecessary questions from those that act but fail to reconcile customer context or fully resolve the request. We release the cases, mock environment, and evaluation harness.

Large Language Models (LLMs) have been used in various industries. However, ensuring their compliance with complex laws and regulatory frameworks remains a great challenge. Existing evaluation paradigms mainly rely on static benchmarks that suffer from three severe limitations: First, the compliance rules being used do not comply with the requirements of Artificial Intelligence (AI) laws and regulations; Second, they only handle apparent, explicit compliance risks, leaving implicit and covert compliance risks undetected; Third, they fail to track the systematic propagation of risks along logical dependency chains or evaluate compliance within nuanced, context-based real-world scenarios. To bridge this critical gap, we introduce EADC, a novel advanced evaluation benchmark of LLMs based on an AI compliance knowledge graph and AI compliance legal experts. By mapping abstract legal rules into structured logical multi-relational graphs, our framework enables automated, evolving agents to distill and synthesize highly sophisticated adversarial scenarios. This compliance benchmark is reviewed and corrected by human AI legal experts throughout the whole process. The resulting dataset (4,435+ QA pairs) provides an extensive, multi-dimensional taxonomy covering critical regulatory frontiers, including bias and discrimination, fairness, personal privacy protection, and values. Crucially, our compliance dataset moves beyond shallow string-matching by incorporating contextual long-horizon interactions and logic-driven hazard chains, capturing deeply embedded compliance anomalies that bypass traditional filters. Experiment evaluations demonstrate that our framework exposes critical regulatory blind spots in state-of-the-art LLMs, offering a rigorous, AI laws and regulations-aligned benchmark to safeguard high-level and deep compliance in the application of LLMs.

Evaluation of large language models (LLMs) for safety, security, and privacy (SSP) relies heavily on static benchmarks, which suffer from score saturation, data contamination, and aggregation artifacts, and fail to capture sensitivity to linguistic variation. As a result, models that perform well on fixed test sets often fail under semantically equivalent rephrasings. We introduce SSP-Bench, a dynamic benchmarking framework that generates evaluation instances on demand while preserving domain consistency. The framework ensures label validity through externally grounded sources, enforces scope via service-specific validation, and calibrates difficulty using a multi-model steering panel. Benchmark construction is formulated as a multi-objective optimization problem over difficulty, separability, novelty, and diversity. Across 24 models and four SSP services, SSP-Bench reveals systematic failures of static evaluation, including near-zero correlation in safety rankings due to construct mixing, strong safety--over-refusal coupling, and hidden within-family regressions. These results show that static benchmarks can misrepresent model behavior, motivating dynamic, deployment-relevant evaluation.

Large language models (LLMs) are increasingly considered for safety-critical engineering, yet their reliability in regulated functional-safety workflows remains underexplored. We introduce SAFARI (Safety-Aware Functional Automotive Risk Inference), the first industrial benchmark for LLM-assisted automotive Hazard Analysis and Risk Assessment (HARA) under ISO 26262. It contains 3,000 de-identified industrial HARA cases and evaluates two coupled tasks: open-ended hazard analysis and standards-grounded risk assessment. To evaluate open-ended HARA artifacts, we propose the first reference-anchored LLM-as-a-judge protocol with high expert correlation. Experiments with nine frontier LLMs show that models often produce plausible hazard narratives but remain weak at ISO 26262 risk classification, with the best ASIL macro-F1 reaching only 0.261. Chain-of-Thought prompting provides limited benefit and often degrades categorical risk assessment. Error analysis further localizes major failures to scenario-critical context omissions during hazard generation and to controllability misjudgments during risk assessment, indicating where expert oversight should be concentrated. The dataset can be obtained from https://github.com/xixi47520-hash/HARA.

The widespread adoption of LLMs has led to escalating content compliance risks. Prior works have contributed to addressing these risks in the English context, downplaying the complexity of Chinese language content. This paper follows China's current AI-Generated content compliance requirements and provides evaluation results on 20 notable LLMs, offering insight into China's regulatory landscape. We design a novel framework to assess the compliance and refusal rates with 2303 questions spanning six distinct dimensions, including 203 self-constructed constitutional questions. The framework employs several judges to generate verdicts independently based on their hierarchical alignment memory. Our findings show that international models also exhibit high levels of compliance despite the use of standard Chinese questions, and the main differences may stem from dimensions closely related to ideological alignment. We establish a regulatory benchmark that enables the global AI community to evaluate both Chinese and non-Chinese LLMs under a unified set of legally grounded compliance requirements.

Large language models (LLMs) are increasingly integrated into vehicle voice assistants. But linking natural-language requests to vehicle functions creates a safety-critical authorization problem. Before executing a command, the system must choose whether to execute, refuse, clarify, require confirmation, defer to manual control, trigger an emergency response, or make no tool call. To our knowledge, prior evaluations do not isolate this pre-action decision across speaker role, authentication status, vehicle state, and tool availability. We introduce a 202-scenario benchmark with Reference Decisions under a seven-class taxonomy. We evaluate two local open-weight models and three API-based LLMs using Decision Alignment and safety-specific error metrics. Alignment ranges from 40.1% for Llama 3.2 3B to 89.1% for Gemini 3.1 Pro Preview. The API-based models score between 83.2% and 89.1%, with no statistically significant differences among them. Even these models produce two to three False Executes among 161 non-execution scenarios, and persistent errors remain in confirmation and manual-control decisions. A controlled Llama 3.2 3B ablation increases alignment to 40.1% under the structured authorization policy, versus 28.2-29.2% under schema-only and generic-safety baselines, but it does not eliminate False Executes. Structured LLM decisions are therefore insufficient as a standalone safety mechanism, and deployment requires an independent enforcement layer that verifies tool permissions and vehicle-state constraints before invoking any vehicle function.

Frontier models are often evaluated on how they respond to safety information once it is already in context. We study an earlier decision point: whether models choose to acquire safety-relevant evidence before acting. We introduce SAFE, a controlled benchmark in which models make deployment decisions with optional evidence that varies in retrieval cost, probability, severity, and presentation. Across GPT-5.5, o3, Claude Opus 4.8, and Claude Sonnet 4.6, we find distinct evidence-acquisition policies: Opus inspects nearly by default, o3 is the most skip-heavy and threshold-sensitive, and GPT-5.5 and Sonnet occupy intermediate regimes. Inspection increases strongly with severity and decreases with retrieval cost, whereas probability has much weaker behavioral influence: increasing the stated likelihood of a problem from 10% to 70% changes inspection by at most 21 percentage points. Despite these differences, Stage 1 rationales are dominated by expected-value reasoning across models. A cost-obligation decomposition further shows that avoidance is driven primarily by retrieval friction and explicit threats to the deployment payoff rather than by the remediation duties created by knowing. Counterfactual interventions reveal a further mismatch between behavior and explanation: evidence framing can strongly change decisions near the inspection boundary while going largely unmentioned, whereas probability is frequently cited despite having little causal influence. These results suggest that deployment-time safety depends not only on how models respond to known risks, but also on whether they acquire the evidence needed to know that acting is safe.

Background: Large language models (LLMs) have achieved physician-comparable performance on multiple-choice medical knowledge examinations, but their capabilities in clinical history taking, urgency assessment, and safety remain insufficiently evaluated. We proposed Japanese Stroke LLM Evaluation, a multi-turn conversational benchmark for stroke care in Japanese, and evaluated LLM performance and safety under practice-oriented conditions. Methods: We created 10 stroke and related-condition cases and evaluated LLMs in multi-turn Japanese conversations. The LLM acted as physician, while a board-certified neurosurgeon acted as simulated patient and evaluator. Each case comprised history-taking and action phases scored using pre-specified criteria. Errors that could directly threaten life were defined as critical mistakes. The safety threshold was at least 80% overall with zero critical mistakes. Eighteen models were evaluated in October 2025 and June 2026. Results: Claude Fable 5 achieved the highest score (87.4%) with zero critical mistakes, followed by Claude Opus 4.7 (80.3%) and GLM-5.2 (75.6%). Two leaders met the safety threshold. Eleven models made 17 critical mistakes, including failure to confirm laboratory results or blood glucose before t-PA, surgery before airway stabilization, omission of cervical vascular evaluation, and t-PA outside its indication. History-taking question count correlated with history-taking score (r = 0.648, p = 0.007). Conclusions: Japanese Stroke LLM Evaluation provides a benchmark for LLM performance under practice-oriented conditions, including a cap on history-taking questions. Cases and evaluations were created by neurosurgical specialists rather than using an LLM-as-judge approach. Performance improved across cloud-based and on-premise models in 2026, with some exceeding the safety threshold. Further evaluation using real-world cases is required.

People increasingly use large language models (LLMs) for mental health support, yet their safety in evolving, high-risk conversations remains poorly characterised. We developed K-Bench, a clinician-calibrated, protected benchmark evaluating 125 model configurations representing 33 base models from 14 providers across a fixed cohort of 200 multi-turn vignettes involving suicide, self-harm, domestic violence, substance misuse, and no-risk presentations. Synthetic patient conversations showed substantial distributional overlap with real human-AI conversations. A frozen GPT-4o judge achieved 94.2% exact agreement with clinician consensus across 6,751 eligible item comparisons from 151 clinician-rated transcripts. Leading models combined strong supportive conversation with combined-risk scores above 95, whereas risk exploration exposed substantial variation among lower-performing configurations. Therapeutic prompting produced configuration-specific gains concentrated among weaker models, while elevated reasoning produced no average improvement. K-Bench combines broader clinical coverage and configuration-scale comparison with a continuously updated public leaderboard whose operational test materials are protected from direct optimisation. The leaderboard is available at www.k-bench.ai.

A structurally valid DeFi workflow can still authorize a costly trade. We introduce DeFiFlowBench, a benchmark of 207 team-authored prompts for natural-language DeFi workflow synthesis. It measures graph coverage, configuration completeness, and declared safety predicates, then tests supported trade configurations on a local EVM. Direct, constrained, and few-shot prompting produce 14-19 unsafe held-out executions per configuration under a fixed 5% price-impact cap. A slippage bound derived from a quote does not prevent the price impact of the order itself. We propose Koan-Safe, which combines a prompt-only intent parser, a replaceable generator, and structural repair with default safety parameters. On 75 held-out workflow prompts, its hybrid variant scores 0.67 on the static safety proxy, compared with 0.33 for the best baseline. Koan-Safe records no unsafe executions on the saved benchmark outputs. A matched-candidate ablation produces 14-17 unsafe executions when enforcement is disabled. Additional tests expose the limits of default injection: permissive existing thresholds can still authorize unsafe trades. A separately evaluated policy cap addresses this failure on a 36-case diagnostic grid. These results support explicit trade protections and execution-based evaluation, while distinguishing declared safety from a general guarantee.

Allowing large language models (LLMs) to retrieve information from a set of trusted documents can increase reliability and reduce hallucination. However, recent work has demonstrated that retrieval-augmented generation (RAG) can have unintended side effects on the overall safety of the generated responses, when prompted for harmful or dangerous content. A clearer understanding of the mechanisms leading to this result is needed, as increasing numbers of end users turn to RAG to incorporate corporate documents and knowledge bases into LLM-based systems. We introduce RAG-Safety-Bench, a benchmark to measure the safety impact of RAG on LLM models. By removing the confounding effect of retriever quality, and cleanly separating the problem into four conditions -- non-RAG, RAG with an oracle document containing the answer to the harmful request, RAG with documents related to the harmful request but without the specific answer, and RAG with random, safe documents -- the benchmark isolates the impacts of different factors in the observed safety degradation. We report results across five open-source LLMs, showing an inverse relationship between benign and unsafe capability, strong evidence that baseline safety guardrails do not lead to downstream safety guarantees in the RAG case, and model-specific support for previous findings that even benign documents can lead to unsafe generation in retrieval-enabled systems.

Large language models (LLMs) may abandon correct positions when users push back, exhibiting a failure mode known as sycophancy. Existing evaluations typically use short, pre-specified conversations and may therefore miss failures that emerge under sustained, adaptive disagreement. We introduce SPINE, a benchmark in which an LLM proxy plays a persistent but mistaken user and adaptively challenges a target model for up to 25 turns. We evaluate four production systems and three Olmo3-7b variants on 100 false-presupposition and 100 unethical-query items. Our experimental results show that collapse rates increase with conversation length for every model, short-horizon protocols underestimate sycophancy and resistance under sustained pressure remains unreliable across current models. By analyzing models with accessible reasoning traces, we surprisingly found that the correct position often remains represented in a reasoning trace when the response concedes, suggesting that the model chooses to please a user and sycophancy is not due to lack of knowledge or ignorance. Ablations show that adaptive LLM proxy exposes more sycophantic collapse than pre-generated scripts. Among all tactics, emotional appeals is the most associated with inducing LLM sycophantic behavior. The code and data are released at https://anonymous.4open.science/r/SPINE

Aligned language models fail under two independent pressures: the structural jailbreak class recently formalized as Involuntary In-Context Learning (IICL), which reframes a harmful request as the final missing cell of a data-labeling task completed by pattern rather than judged as content; and the erosion of safety alignment outside English. A natural hypothesis is that these compound. We test it directly. Using a deterministic IICL operator and a StrongREJECT-style rubric judge, we red-team two Google Gemini models on two benchmarks, a 30 general-harm behaviours from HarmBench and 30 financial-abuse behaviours from FinProof, each under a single-shot baseline and under IICL in four languages (English, Spanish, Hindi, Arabic). First, IICL generalizes to a second provider and is worse in finance: it lifts attack success from <=6.7% to 80-90% on HarmBench and 97-100% on FinProof, an order of magnitude above the <=24% its introducing study reported on OpenAI's GPT-5.4. Second, against the hypothesis, forcing the IICL output into a non-English language does not stack the two weaknesses, it attenuates the attack. Eleven of twelve non-English conditions score below their English baseline (sign test, p~0.003), the lone exception a ceiling tie near 100%; on the stronger model's financial set Arabic collapses from 100% to 33%. We attribute this to a relevance curse: once structure has unlocked compliance, the models produce lower-quality harmful content in lower-resource languages, which a substance-grading judge scores as partial. The pattern replicates under an independent non-Google judge (Cohen's kappa=0.86, 377 paired verdicts), and 76.6% of non-English responses were verified in-language. Jailbreak vulnerabilities are therefore not additive; the dominant residual risk is the English structural attack, most acute for financial abuse, not a multilingual one.

Automated red-team attacks and blue-team defenses for large language models (LLMs) are advancing quickly. However, attackers and defenders are built and tested in isolation, and the resulting scores are hard to trust. To tackle this, we present ACEA (Adversarial Co-Evolution Arena), a platform that connects a pluggable red-team adapter and a pluggable blue-team adapter to a shared target LLM and scores their attack and defense rates with an LLM judge. ACEA contributes four components. First, a pluggable, model-agnostic arena. Any red or blue project connects over a minimal HTTP protocol, which we call the ACEA Standard Adapter Protocol (ASAP). It can be written in any language, and a project that exposes nothing but the protocol is a full participant. Second, an evaluation methodology built for adversarial rounds. Seeding the target with canonical secrets gives verifiable ground truth that separates real leakage from hallucination. We also send each attack to the target even when the defense blocks it, which measures the attack's raw potency independently of whether it was stopped. Together these yield a per-round decomposition of attack strength and defense effectiveness. Third, a real-time, game-style visualization with a detailed end-of-battle report that localizes each failure. The evaluation thus becomes an actionable signal for improving a red or blue project. Fourth, an optional in-context improvement loop that turns each round's outcome into advisory hints for the next. An adapter can then adapt across rounds without keeping state, provided it reads the hints. We describe the design of ACEA and the metrics through which red and blue teams are scored head to head.

Automatic safety judges -- systems such as Llama Guard or a GPT-4o grading prompt that decide whether a model's reply is harmful -- produce the numbers behind almost every reported jailbreak success rate, defense evaluation, and safety leaderboard. We ask whether these judges grade what a reply contains or how it sounds. We keep a reply's content fixed and add content-invariant style wrappers: fixed strings placed before or after the reply that change only its tone (an educational disclaimer, a fake safety "reasoning" block, a token refusal followed by the unchanged harmful body), or, on harmless refusals, framing that merely sounds dangerous. The body is preserved byte-for-byte, so a faithful judge must return the same verdict, and any flip is an error of the judge, not a change in safety. Over 600 JailbreakBench replies x up to 7 forms x 8 judges, we measure flip rates with paired significance tests and measured noise floors. Findings are precise rather than universal: most judges barely move, but specific judges harbor cheaply exploitable blind spots. A token-refusal wrapper flips 19.9% of GPT-4o-mini's correct "unsafe" verdicts (noise floor 0.5%; 18.2% under majority-of-three re-scoring) yet moves Claude only 0.4%. The deployed Llama Guard 4 is deterministically gamed: an "educational course" framing flips 12.3% of its harmful verdicts to safe. A second deployed guard (gpt-oss-safeguard-20b) is immune, and rewriting only the grading prompt (StrongREJECT-style) cuts the attack tenfold on the identical model -- the vulnerability lives in the judge, not the content. A two-annotator human validation confirms 100% content invariance and 90% of flips as judge errors (kappa 0.95-1.0), and a bootstrap shows the underlying model ranking is already unstable to sampling alone. We release the dataset, wrappers, code, and per-verdict labels.

Multimodal Large Language Models (MLLMs) show strong progress on vision-language tasks, yet their reliability in safety-critical settings remains underexplored. Fire-smoke understanding is central to public safety and disaster response, but most existing benchmarks lack diverse real-world scenarios and context-aware evaluation. We introduce SAFIRE, a large-scale benchmark for fire-smoke understanding in MLLMs, comprising 83K captioned images from 20 scenarios and 193K multiple-choice VQA (MCVQA) generated from a 9.7K-image subset, spanning 10 evaluation dimensions from basic perception to higher-order reasoning. A GPT-5.4-assisted multi-stage verification pipeline with MLLM majority voting ensures annotation quality. Evaluating ten open-source MLLMs (8B-38B) yields an average accuracy of 61.9%, exposing major gaps in safety-critical reasoning. We further show that adapting vision encoders with only 7% of our domain-specific data boosts fire-scene classification accuracy from 20.1% to 64.5%, indicating that carefully curated data can yield substantial gains even when data volume is limited. All datasets, models, and code are available at https://risys-lab.github.io/SAFIRE/.

Several properties safety monitors are asked to certify, among them cross-tenant noninterference, sandbagging and evaluation awareness, are 2-safety hyperproperties, witnessed only by two executions. The standard consequence is a binary impossibility: one trace cannot decide them. We replace the binary with a measurement. A tight bound puts the balanced accuracy of any single-trace monitor at $\tfrac12+\tfrac12\,TV(P_0,P_1)$, turning undecidability into a graded detectability frontier and defining an oversight gap: a monitor's shortfall below it. On a leak family with closed-form $TV$, nine LLM monitors are optimal at $TV=0$ but capture little signal as $TV$ grows; at $TV=1$, where a 20-line membership check scores $100\%$, they average $60.9\%$. That shortfall is mostly not capability: naming what to check closes $61\%$ of it while leaving the $TV=0$ control at chance. The same split runs through a $2{\times}2$ factorial: an imagined second run leaves monitors at chance ($50.4\%$) while the same rule on an executed second run reaches $90.0\%$, and a stored oracle without a comparison procedure yields only $68.2\%$. Information and procedure are each necessary and neither is capability. Under nondeterminism, replay tracks a closed-form $k$-replay curve only under the right projection, and a projection frontier shows the resulting dilemma is forced: narrow misses $98.6\%$ of off-channel leaks, broad flags $75.7\%$ of clean traffic, and attainable accuracy decays like $1/(qm)$ in the benign-variation rate and the channel count. Finally, two frontier LLM judges certified an earlier version of our own benchmark as sound while a sign test found a directional bias ($p=2.7\times10^{-5}$) that invalidated three of our findings. Construction validity for hyperproperty benchmarks should be proved mechanically, not audited by models.

Evaluating large language models (LLMs) in safety-critical, physics-governed environments requires more than accuracy-based metrics, because predictions that are numerically close to the ground truth can still violate operational constraints, combine fields in physically inconsistent ways, or fail to produce usable structured outputs. Existing evaluation protocols do not measure these failure modes reliably. We propose FLY-EVAL++, an evidence-driven evaluation protocol that combines deterministic verification of protocol compliance, physical feasibility, and safety constraints with fixed rubric-guided aggregation into interpretable multi-dimensional scores. We instantiate FLY-EVAL++ for Flight Trajectory and Attitude Prediction (FTAP) by extending the PilotBench setting with history-conditioned and multi-step prediction tasks. Across 66 LLMs, safety compliance is the most discriminative dimension of model behavior: models with comparable predictive performance differ by more than 28 points in safety score, and we observe recurrent failures including safety violations under physically plausible predictions and instability in multi-step rollouts. These results show that evaluation in safety-critical domains should measure constraint satisfaction and structured validity explicitly rather than rely on accuracy-centric reporting alone.

Large language models (LLMs) demonstrate strong performance on standard content moderation benchmarks. However, these benchmarks often aggregate multiple moderation criteria into a single label, making it unclear whether models can disentangle them and reliably apply each criterion when making decisions. To study whether LLMs exhibit criterion-conditioned behaviour, we introduce Diagnostic Evaluation of COntent (DECO), a criterion-independent factorisation of content that enables controlled, criterion-level evaluation. We also introduce pairwise evaluation to compare model outputs across different criteria for the same input. Across four moderation datasets and four LLMs, we find that strong benchmark performance can hide substantial failures at the criterion level. Models struggle most when correct decisions depend not on overall harmfulness, but on the specific aspect of the content that the criterion requires them to assess. Our results highlight a key limitation of current content moderation benchmarks: strong performance on aggregated labels does not provide sufficient evidence that LLMs can reliably evaluate content with respect to individual moderation criteria. These findings call for the development of evaluation methods that explicitly measure criterion-conditioned behaviour.

Large language models (LLMs) are increasingly used in multilingual settings, yet their safety is still evaluated primarily in English. This limits our understanding of how alignment failures manifest in low-resource and culturally diverse languages. We introduce IndicSafeEval, a persuasion-based jailbreak evaluation framework for Indian languages. Our benchmark combines ten safety critical content categories with six human-like persuasive strategies across four different Indian languages, such as Hindi, Bengali, Marathi and Punjabi, resulting in 7,200 adversarial prompts. We conduct a systematic black-box evaluation of several open-source LLMs to examine how their safety behaviour varies across languages, persuasion strategies, and risk categories. Our analysis shows that the model does not behave equally safely across all languages and prompt styles. Instead, safety performance depends strongly on both the languages used and the way a request is phrased using persuasive cues. We further observe that different risk categories exhibit different levels of vulnerability, with some types of harmful content being significantly more susceptible to persuasion-based jailbreaks than others. These findings reveal important limitations of current safety evaluations, which are largely English-centric, and underscore the need for multilingual and persuasion-aware benchmarking frameworks to more accurately assess real-world LLM safety. Our implementation is available at https://github.com/MonSaikat/IndicSafeEval. Warning: this paper contains example data that may be offensive or harmful.

Fragmented safety evaluation undermines the governance of dangerous AI capabilities. We present a modular framework that evaluates each model through three orthogonal pipelines---Knowledge ($K$), Defense ($D$), and Harm ($H$)---under a unified protocol, aggregating results into a standardized dangerous-capability profile $φ$. Pluggable modules supply scenario seeds, knowledge banks, hazard queries, and judge rubrics, while the core evaluation engine remains unchanged across domains; the CB evaluation is complemented by a cyber pilot demonstrating protocol transfer. Instantiating the framework with a chemical-biological (CB) module, we evaluate 12 commercial LLMs from four families. Our first contribution is a horizontal comparison of dangerous capability across models and model families: the three dimensions expose sharply divergent profiles---models with comparable knowledge differ in refusal resilience, and strong defenders do not generate less harmful content when they do comply---while family-level patterns further separate Claude, DeepSeek, and GPT models. The second is a temporal analysis of capability evolution: tracking $K$, $D$, and $H$ against model release dates reveals that dangerous capability has not monotonically declined; newer models deepen knowledge while only partially improving defense, showing that scaling and alignment progress do not uniformly translate into safety. Reliability is established via cross-judge consistency (bootstrap $ρ> 0.79$, 4 of 5 judges) and pipeline orthogonality ($K$--$D$--$H$ inter-correlations $ρ\in [0.32, 0.52]$).