Safety Evaluation

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.

Large language model (LLM) agents increasingly operate over long-horizon interactions involving tool use, persistent state, evolving authorization, and external environment feedback. In such settings, safety failures may emerge only after multiple turns, yet existing evaluations often reduce agent behavior to task or attack success, obscuring whether an agent acts, refuses, or remains appropriately calibrated as the interaction evolves. We introduce Blindspot, a benchmark for trajectory-level safety calibration of long-horizon tool-using agents. Blindspot evaluates complete user-agent-environment trajectories through adaptive adversarial interaction, stateful tool execution, and execution-grounded adjudication. Its current instantiation contains 22 attack families and 35 scenarios across seven domains, yielding more than 2,500 long-horizon trajectories with an average interaction length of 14.7 turns. Each trajectory is assigned one of five outcomes: Safe Completion, Correct Refusal, Unsafe Completion, Over-Refusal, or Indeterminate. Unlike fixed attack datasets, Blindspot is an extensible live-simulation framework in which attacks, scenarios, tools, policies, domains, and agent configurations can be added without redesigning the evaluation pipeline. We evaluate 13 proprietary and open-weight LLMs using eight metrics covering unsafe completion, appropriate refusal, benign utility, over-refusal, repeated-run robustness, and post-refusal failure. Preliminary results reveal substantial differences in safety-utility calibration across models and show that failures can emerge only after several initially safe interaction steps. These findings motivate treating agent safety as a trajectory-level property rather than a single-turn or binary success criterion.

Safe reinforcement learning (RL) is commonly formalized as a Constrained Markov Decision Process (CMDP), in which an agent maximizes expected reward while keeping its expected cumulative cost below a specified safety bound. Existing safe RL benchmarks predominantly report whether an algorithm is safe on average, following this expectation-based guarantee. We argue that this convention is insufficient to reliably characterize an algorithm's true safety: it fails to capture how often and how severely the safety bound is violated, whether this holds consistently across tasks and safety bounds, and whether training-time behavior is representative of behavior of the final converged policy. Therefore, we introduce (i) evaluation metrics for safe RL that address each of these concerns and in addition allow for aggregation across tasks and safety bounds. We furthermore define (ii) a safety tier system to systematically categorize and compare algorithms in terms of safety and reliability at both training and for a final policy. Using this framework, we provide (iii) an empirical safety evaluation across multiple safety navigation tasks. Our results show that aggregate metrics, distributional reporting, and task- and safety bound-specific results each reveal information the other metrics cannot. We therefore recommend reporting all three jointly, rather than compressing this information into a single value, as is common practice. We provide SafeRLEval, an open-source evaluation suite to support the reliable characterization of safety in future safe RL research.

Increasing safety is the primary objective of automated vehicles. Achieving this goal requires reliable safety metrics that incorporate safety-relevant factors such as object type, velocity, and criticality. A key capability of such metrics is the distinction between critical and non-critical objects, which is addressed through criticality or relevance estimation. Existing criticality metrics are typically designed for specific scenarios and primarily focus on vehicle-to-vehicle interactions. In this paper, we therefore propose a novel criticality metric tailored to vulnerable road users (VRUs), which require special consideration due to their less predictable motion behavior. Furthermore, to avoid the complexity introduced by scenario-specific metrics, we introduce a scenario-independent criticality prediction framework applicable to all traffic participant classes. The effectiveness of both the proposed VRU-centric criticality metric and the criticality prediction framework is evaluated using the DeepAccident dataset, which contains a diverse set of safety-critical traffic scenarios. The proposed VRU-centric criticality metric improves pedestrian criticality classification performance by up to 50 %. In addition, the proposed criticality prediction framework outperforms state-of-the-art metrics by 275 %, achieving an F1-score of 0.96 and enabling scenario-independent criticality assessment across all object classes. These results demonstrate the strong potential of the proposed approaches to enhance criticality assessment for safety evaluation in automated driving systems.

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.

Automating alignment research may accelerate progress toward aligned AI, but whether it does is hard to measure. Luckily, many alignment failures, such as deception, sycophancy, and jailbreaks, are already measurable by public benchmarks. We study whether automated alignment researchers (AARs) can post-train to mitigate alignment failures by proposing training methods and data to simultaneously optimize multiple safety benchmarks, while largely preserving general capability. Across 10 alignment failures, the strongest AAR methods significantly reduce the targeted alignment failures and generalize to a held-out benchmark, multi-turn behavioral audits, and models up to 4.7x larger than the target model. As a human baseline, 28 experienced researchers receive up to eight hours to develop one-shot methods for the same benchmarks, but their methods underperform the best AAR methods. Using human ideas as the AARs' initial research direction does not improve performance, suggesting current AARs may not need guidance from experienced researchers. These results suggest that automating alignment research on well-characterized failures may be practical in the near term.

Can language models be trusted in safety- critical operations? In such settings, strong per- formance on semantic metrics does not guaran- tee operational reliability: a misread altitude, a dropped execution condition, or a confused call- sign may score well under standard F1 yet carry sharply asymmetric operational consequences. We study this problem in air traffic control (ATC), where controller-pilot communication demands near-zero error tolerance, and use consequence-aware evaluation to test whether semantic scores misstate operational reliabil- ity. The framework is instantiated in a con- trolled diagnostic ATC benchmark grounded in aviation standards and feedback from 40 air traffic controllers across three countries. Evaluating 8 models, we uncover a system- atic semantic-safety gap: conventional scores give substantially higher performance estimates than consequence-aware evaluation, even for models that appear reliable under standard met- rics. Risk-aware fine-tuning narrows but does not close this gap, showing that consequence- aware evaluation is a necessary complement to standard NLP metrics before any real safety- critical deployment claim

Existing safety evaluation datasets for large language models (LLMs) predominantly focus on English and Western contexts, often overlooking the linguistic diversity and culturally grounded safety risks present in other languages. To address this gap, we introduce SurakshaEval, a novel safety benchmark composed of human-written prompts spanning real-world scenarios, explicitly designed for ten major Indian languages - Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Punjabi, Tamil, and Telugu, along with English. SurakshaEval includes both generic prompts common across India and region- and language-specific prompts that capture localized sociocultural sensitivities. We benchmark a broad range of state-of-the-art LLMs on SurakshaEval, establish baseline safety performance, and identify recurring failure modes, including over-refusal, missed detection of implicit bias, and insufficient contextual awareness in regionally sensitive settings. Our results show that even strong multilingual LLMs struggle to reliably meet nuanced safety requirements when operating in Indic languages, particularly in native scripts. These findings highlight the urgent need for safety evaluation frameworks that incorporate region-specific data and structured assessment protocols, enabling the development and deployment of AI systems that operate securely, ethically, and in alignment with diverse societal values. Our code and data are available at https://github.com/debobanerjee/SurakshaEval. Warning: This paper contains text that may be offensive or unsafe.

Large language model (LLM) benchmark evaluations are routinely used to support claims about model safety, reliability, and deployment readiness. Yet most evaluations rely on a single access modality (model APIs), perform a single run per prompt, and report accuracy as the primary outcome metric, without accounting for conditions such as web search that may have effects on model behavior in deployment. We audit these assumptions for one of the most widely-used LLMs, comparing two modalities, ChatGPT's chat UI and OpenAI's API, with and without web search enabled. We use a stratified total sample of 401 prompts from two popular benchmarks, BBQ and SafetyBench, collecting 4,812 total responses across three repeated runs per prompt. Beyond standard performance measures, we evaluate model output dimensions including response consistency, response text similarity, citation grounding, and abstention behavior. For instance, chat UI responses were less accurate than API responses on both benchmarks with search disabled. Enabling web search reduced accuracy by up to 8 percentage points, and even reversed the direction of modality performance trends for one benchmark. Repeated runs of the same prompt produced inconsistent responses in up to 21% of prompts. The two modalities also grounded answers in different citations, and abstention behavior was also inconsistent across both modalities. These results illustrate that, even within a model family, reporting only simple accuracy metrics can obscure important forms of model behavioral variation relevant to AI safety assessments. We argue that AI safety evaluations should systematically account for modality, multi-run consistency, search conditions, and response-level behaviors to better reflect how deployed AI systems behave in practice.

Defensive driving scores are useful only when they preserve distinctions between policies that observe surrounding actors and those that do not. Re-simulation benchmarks may use reference-conditioned forgiveness, under which an agent receives credit when the logged human reference fails a compliance channel. When agent and reference share an unstable rollout transformation, this rule can propagate shared reference failures into broad compliance credit. We audit this risk in NAVSIM v2.2 original scene single-stage scoring. Under the affected documented-stack condition on the audited numerical backend, the route-blind Ignore-All probe and a route-aware actor-blind probe outrank human replay and PDM-Closed over the complete 12,146-token navtest split. A fresh installation following the public specification reproduces rollout divergence on a fixed 32-token diagnostic set. A same-source dependency stack control and an exact-input diagnostic isolate dependency-sensitive numerical behavior in the shared velocity refit. On a 450-token control pool, replacing only the solver eliminates rollout divergence and restores blind-last ordering while keeping forgiveness enabled. Thus, the numerical instability is the direct trigger. Reference-conditioned forgiveness propagates the resulting shared reference failures into compliance credit. We contribute an audit protocol requiring score basis and stack disclosure, blind probes, overwrite reporting, and rollout stability tests before using such scores for defensive driving claims.

Task-specific fine-tuning can improve the performance of large language models (LLMs) on downstream tasks. However, our study reveals that task-specific fine-tuning can also weaken the safety guardrails of aligned LLMs. A widely adopted strategy for preserving safety during fine-tuning is to incorporate safety data. Although previous studies have shown that randomly mixing safety data can alleviate safety degradation, the underlying principle determining why some safety examples are more effective than others still remains unclear. In this paper, we propose DataRx, a missingness-aware sampling method for selecting safety-critical examples. DataRx is based on the hypothesis that a safety sample is more effective when the selected examples provide safety signals that fill the missing parts of LLMs' safety capabilities. DataRx's key insight is leveraging high-dimensional hidden representations rather than discrete tokens to quantify the safety signal gap between the target model's native response and the safety reference response. The results show that, with only 1% additional safety samples from BeaverTails, DataRx reduces the average attack success rate of Llama3-8B-Instruct across seven downstream tasks from 59.23% under random sampling to 13.70%. In addition, DataRx can be combined with the existing safety data synthesis method to further enhance safety defenses during fine-tuning. We hope that DataRx will inspire more data-centric defense research.

This white paper presents ADMITBench, a reference framework for evaluating industrial LLM advisories at the level of the proposed action. The framework implements a versioned, safety-governed evaluation contract that checks whether a recommendation is supported by the available evidence, permitted under the stated authority and procedure, and acceptable under the plant-specific consequence checks encoded in the selected evaluation profile. In this report, \emph{safety-governed} means that eligibility is determined through explicit, non-compensatory checks derived from a versioned plant profile; it does not mean that the evaluator, model, or plant has been safety-certified. Release 0.1.0 is a public reference implementation for technical and research evaluation, not an authorisation for physical execution.

Applying a valid medication-safety rule when its patient-specific conditions are not met can produce an incorrect decision. Existing medical evaluations largely use isolated and fixed scenarios. A model may therefore answer correctly by recalling a drug-risk association without showing that it used patient information to decide whether the rule applies. To address this gap, we introduce MedPIC-Bench, a benchmark of source-verifiable recommendations and expert-validated questions for patient-specific medication-safety reasoning. It combines guideline-following questions with paired counterfactual questions in which a controlled change in patient information changes whether a rule applies. The benchmark contains 467 questions annotated along six clinical and reasoning dimensions. Across 28 medical-specific, general, and proprietary LLMs, every model performs worse on counterfactual questions, with mean accuracy falling from 63.6% to 45.1%. Models perform well when an explicit patient attribute directly signals a familiar contraindication, but struggle when patient information must narrow or withdraw a safety warning. Model rationales often acknowledge the changed patient information, yet the final answers retain the previous safety judgment. This vulnerability persists among medical-specific LLMs, whose average CF performance trails that of general LLMs. MedPIC-Bench therefore makes conditional rule application measurable and highlights the limitations of static medication-safety accuracy for assessing patient-specific reliability.

Design campaigns in chemistry, materials science, and machine learning share a bottleneck: determining how good a candidate truly is requires an expensive evaluation - an experiment, a first-principles simulation, or a full training run. Machine-learning surrogates that predict these outcomes are increasingly used not only to propose candidates but to grade them, and even to feed their own predictions back into the search as though they were measurements. Through mathematical analysis validated on three exhaustively ground-truthed design tasks, we establish when this practice is safe, what any certificate of safety must cost, and when the substitution provably pays. Predictive accuracy cannot anchor trust: near-perfect R^2 is compatible with worst-possible selections, and screening N candidates inflates the over-prediction at the selected candidate by a quantifiable "selection tax" with matching upper and lower bounds. Safety follows instead from an architectural rule - predictions may propose and train without restriction, but every certified conclusion must rest on true evaluations - which is sufficient with no assumptions on the surrogate, and necessary, since admitting predictions into certification with the standing of measurements opens a deterministic self-confirmation failure mode. We derive the minimal criterion under which a model may act as an oracle (rank preservation, not accuracy), show that trust must be purchased through selection-aware audits that are optimal in query complexity, and prove a dichotomy fixing when audited surrogates cut certified evaluation cost. Across 432 surrogate fits over six task-regime conditions, the audit statistic tracks deployed search performance at Spearman rank correlation 0.80-0.99, while the rank correlation of R^2 with deployed regret falls as low as 0.33; audited screening reduces certified oracle cost by a measured factor of 25.

AI agents operate in persistent environments where early state changes can influence decisions far into the future. Unlike conventional language-model interactions, agent behavior is mediated through a shared state that is repeatedly modified and reused across long-horizon workflows. Current safety benchmarks often fail to capture these cumulative risks because they focus on short, static tasks. To address these limitations, we introduce OpenART, an open-ended arena for scalable agent red teaming through environment evolution. OpenART provides over 10,000 validated stateful scenarios across 50 domains, drawing from a pool of more than 500,000 tools and skills. These tasks require a median of 97 tool calls and enable unified evaluation across 75 different agent-model configurations. To systematically explore these evolving attack surfaces, we propose the Evolutionary Markov Hypergraph Attack (EMHA). EMHA is a black-box policy that performs feedback-driven environment evolution by coordinating authorized state transitions without requiring parameter updates. Throughout the evaluation, task objectives remain fixed while only the environment state changes. Across all configurations, EMHA achieves a pooled Attack Success Rate (ASR) of 85.0%. Its advantage over instruction-only evolution increases from approximately 2% on simple environments to over 17% on the most complex ones, demonstrating that environment evolution increasingly exposes safety failures as task complexity grows. Furthermore, our analysis shows that the specific runtime implementation of an agent explains a significant portion of safety variation beyond the underlying model's capabilities. These results establish OpenART as a scalable foundation for studying agent safety in complex, evolving environments.

In this paper we present an active, constantly updated AI benchmark which measures the integrity of frontier language models against being co-opted for use by authoritarian state "information operations": intentional, coordinated activities by one state to influence public opinion and information ecosystems in another state. These information operations are a well documented, persistent threat against contemporary democracy. Our benchmark is based on real examples from over 2,100 information operations drawn from a live monitoring pipeline which tracks online information assets with links to authoritarian regimes. Alongside this paper, we also release a companion website that updates the benchmark weekly with new claims. The dynamic nature of this public facing benchmark makes it resistant to saturation. In the benchmark, we test 17 models from 8 providers across four prompt framings. We find that most models can be co-opted for information operations at least some of the time. Integrity scores, defined as the share of judged responses in which the model neither preserved nor amplified the claim, range from 9.3% to 91%, an 81.7-percentage-point spread not explained by model size. Models approach participation in information operations in a variety of ways. Some models fabricate details and produce output more harmful than the original input claim; others make claims less harmful even while complying and producing some output. Fact-checking rates vary from 3.2% to 80.8%. Integrity against information operations is at least partly related to refusal to produce content even for benign claims, illustrating the challenge of balancing model usability with safety. Overall, our results show the potential for contemporary information operations to be substantially aided by frontier AI models.

Agent-safety benchmarks measure different behaviors, and their scores get quoted interchangeably as an agent's safety. We treat four of them (R-Judge, InjecAgent, AgentHarm, AgentDojo) as measurements to be validated, running each under its official implementation and author-provided scorer on up to 22 models, with MMLU and GPQA measured by us under one protocol as a capability composite. The metric is the first problem. On any binary trace-judgment benchmark scored by $F_1$, an ``always positive'' policy attains $F_1 = 2π/(1+π)$; on R-Judge that is $0.690$, above five of the 21 models that actually discriminate. The three broad-coverage benchmarks then rank the same 18 models differently, and the trade-off behind that disagreement is a small-panel artifact: R-Judge specificity against AgentHarm safety correlates $-0.64$ at $n{=}7$ and $+0.02$ at $n{=}18$, and a quarter of random size-7 subsets reach $|ρ| \geq 0.5$ around that near-zero value. Held-out validity turns on which outcome you pick. Capability predicts task success ($ρ{=}{+}0.60$) but correlates negatively with misalignment safety ($ρ{=}{-}0.44$, $n{=}21$). On their paired $n{=}20$ panel, the corresponding contrast is $Δ{=}{-}1.00$ (95% CI $[-1.48, -0.49]$, $p<0.001$), and it survives leave-one-organization-out and organization-clustered bootstrap analyses. On an expanded 41-model panel, the misalignment correlation weakens to $-0.16$ (95% CI $[-0.54, +0.22]$) and jailbreak strengthens to $+0.34$, though neither change is significant. \mbox{AgentHarm} shows the strongest held-out association, $ρ{=}{+}0.72$ with three-template jailbreak safety after controlling capability. But both instruments score harmful compliance, so this is evidence of convergent validity rather than general safety. Naming the benchmark, metric, target behavior, and model panel is the minimum a safety claim needs.

Construction-safety models must handle concrete deployment risks, such as a worker standing near a scaffold edge without guardrails, rather than only recognize common objects in curated images. Yet real inspection archives are redundant, long-tailed, and collected across changing sites and months. We introduce SafeBuild-Bench, a metadata-driven benchmark for evaluating multimodal large language models on construction safety under realistic temporal and site variation. It is mined from 100K+ industrial image-text records and contains 3,314 task instances from over 3,000 expert-verified images, covering multiple-choice hazard identification and free-form hazard description. To make expert verification scalable, we develop GEMS, a graph-enhanced multimodal selection pipeline that combines a proxy-model confusion signal with graph-based diversity to identify informative candidates from redundant streams. On public instruction-tuning data, GEMS-selected subsets preserve robustness-oriented performance under small data budgets. On SafeBuild-Bench, current MLLMs remain far from reliable construction-safety understanding, with the best overall score near 60. We release the benchmark, evaluation scripts, and GEMS codebase at https://github.com/safebuild/gems.

Almost every large language model that reaches a broad audience is quantized: trained in full precision, then compressed for efficiency. This step is assumed harmless and its safety is rarely re-checked. We find its principal side effect is increased bias that standard safety evaluation misses. Holding the model, its training, and the prompts fixed, a quantized model still refuses harmful requests, still avoids over-refusing benign prompts, and still selects the unbiased multiple-choice answer. Yet asked an open-ended question, the same model volunteers stereotypes in all eight languages we probe, in roughly one in four open-ended answers under an independent judge (~24% to ~27% across the compression ladder): it passes every standard check and still reaches users measurably more biased. The selective gap is a robust finding; whether open-ended bias further increases with compression is less certain, sensitive to the judge that scores it. We address both with \textbf{QuantiBias}, a benchmark that pairs a generative, multilingual stereotype probe with the refusal and multiple-choice controls that isolate open-ended generation, contrasts each build with and without reasoning, and rates the content severity of what it generates. Across two backbone models (Qwen and Gemma), a five-family screen, and eight benchmarks, quantizers allocate their extra precision by capability data that carries no bias-prevention signal, and reasoning before answering roughly halves the effect on some families while doing nothing on others. A quantized build must be re-evaluated for open-ended bias, not only on the short-form safeguards it already passes.

LLM-based agents leverage third-party skills to extend their capabilities in open-world scenarios. However, third-party skills can introduce extra security vulnerabilities, as seemingly harmless skills can contain latent safety risks that only emerge during actual execution. In this work, we conduct a systematic investigation into how well current agent systems recognize and avoid such risks. To support quantitative and qualitative evaluation, we construct OpenSkillRisk, a dedicated safety benchmark containing 263 risky skills collected from public skill marketplaces. We classify these skills into seven categories based on their threat types and pair each skill with a standardized user task and a corresponding sandbox for controlled evaluation. Distinct from prior benchmarks, OpenSkillRisk not only covers more realistic and diverse unsafe scenarios, but also provides a fine-grained analysis to diagnose the behavioral patterns of agents in such scenarios. We conduct comprehensive experiments covering three mainstream CLI agent frameworks and thirteen state-of-the-art LLMs. Experimental results show that no tested system handles risky skills reliably: even the safest configurations still execute unsafe actions in about 17% of cases. Context-dependent and system-level risks are especially difficult for current agent systems to avoid. Our behavioral analysis reveals three recurring failure patterns: agents may fail to recognize the risk, recognize it but fail to intervene before acting, or follow skill instructions beyond the user's intended scope. These findings highlight the need to improve both risk reasoning in LLMs and execution control in agent frameworks.

Security rating platforms summarize externally observable cyber exposure and are expected to help organizations prioritize remediation. A platform may want to tell an organization how a candidate remediation action would affect its score, but repeatedly exposing exact score responses can reveal information about the hidden scoring engine. We propose a surrogate based approach for remediation score impact prediction that is designed to respect this opacity constraint. The surrogate predicts scores from organization configurations while explicitly representing checkpoint (i.e., a security check) applicability and the observed checkpoint set. A main challenge is that such predictions are not uniformly reliable: they depend on the amount and structure of the observable checkpoint evidence available for a given configuration. To address this, the approach combines applicability-aware surrogate construction, sensitivity analysis under controlled checkpoint restriction, a reliability layer for identifying unstable predictions, and score-impact prediction for supported remediation actions. Explicit modeling of checkpoint applicability is central throughout: it improves score prediction and provides the feature basis used by the reliability layer to identify unstable cases. We evaluate the approach on a real-world dataset of 5,188 organization configurations from a commercial security rating platform. The results show that the applicability-aware surrogate improves score prediction over simpler feature representations. For remediation, the surrogate predicts the score impact of supported actions, while the reliability layer helps identify cases in which these predicted impacts should be interpreted cautiously.

Vision-language models (VLMs) are increasingly used as the reasoning backbone of embodied agents, enabling robots to interpret visual scenes, follow language instructions, and plan multi-step actions. In household environments, however, safety depends not only on recognizing objects, but also on how actions change the physical scene over time. Existing embodied safety evaluations largely focus on static risk recognition, unsafe instruction refusal, or final-state task completion. As a result, process-level safety failures induced by spatial relations such as support, containment, and proximity remain insufficiently studied. To address this gap, we introduce SAFERELBENCH, a spatial-relation-aware safety benchmark with 507 executable evaluation samples, including 248 spatial-relation samples and 259 non-spatial control samples. Using SAFERELBENCH to evaluate seven open- and closed-source VLM-driven embodied agents, we find a substantial gap between task success and process-level safety compliance: models often complete the requested task while violating process-level safety constraints. Unlike prior benchmarks, SAFERELBENCH explicitly tests whether agents satisfy safety conditions before risk-prone actions, making spatial relations a core dimension in embodied safety assessment. More broadly, our results show that safe embodied intelligence requires not only stronger perception and planning, but also reliable reasoning about how object relations shape risk during interaction.

Jailbreak-robustness research typically evaluates safety through generated responses using an LLM-as-judge approach. Such evaluations, however, are sensitive to the benchmark's grading procedure and capture only observed behavior on a given set of attacks, without directly revealing the hidden fragility of the underlying safety mechanisms. This work proposes JADR (Jacobian Assessment of Danger Recognition), a protocol that measures a model's internal representation through Jacobian space (J-space, a recently proposed workspace of verbalizable concepts) before the first response token is generated. For every prompt and layer we record the top-k J-space tokens; these are grouped into six behavioral scenario axes and compared between a danger sample based on StrongREJECT and a safe control drawn from XSTest and OKTest. The method does not call on an external judge model: the computation runs entirely locally, on the activations of the model under evaluation, which lets us compare both different models against each other and modifications of a single model -- quantization and fine-tuning in particular -- on the same terms. The final comparison rests on the proposed SafetyAUC metric, complemented with bootstrap confidence intervals. The protocol is applied to six models (Qwen3-1.7B, Qwen3-4B, Qwen3-8B, Qwen3-Uncensored-4B, Qwen3-SafeRL-4B, Gemma 2 9B) across three weight-representation regimes -- BF16, INT8, and INT4 -- and checked against an independent behavioral evaluation with the StrongREJECT grader. The metric separates models with a strong versus a weak internal safety mechanism with statistical significance and captures substantively different effects across quantization regimes.

Many agent-safety evaluation results are not yet load-bearing evidence: identical nominal outcomes (task success, attack success, monitor scores) may sit atop materially different evidence regimes. No vendor-neutral, runnable instrument scores reconstructability as an evaluation-validity metric: whether captured evidence can reconstruct the decision a claim depends on. This paper introduces a property-level reconstructability metric over eight decision-property classes and a cross-harness adapter emitting per-decision Evidence Sufficiency Cards backing a per-run monitor-coverage release check. It specifies a counterfactual-replay intervention protocol, implements its replayability-precondition probe, and defines a claim-evidence overclaim gap. On public and bundled traces, without new model runs, twelve-field sufficiency spans 0.458-0.833 across four inputs sharing a surface reading; replay preconditions are unmet in every scored trace. In a synthetic release-gate pair, the sufficiency gate blocks the raw variant (0.542) and passes the instrumented (0.667). Safety-evaluation claims should travel with their reconstructability vector; a reproducibility package regenerates every reported number.

LLM agents increasingly perform autonomous actions through external tools, leading to complex and evolving safety risks. However, existing safety testing targets expert-designed safety violations, and the corresponding outcomes are evaluated by hard-coded rules, making them costly to extend as agents evolve. To this end, we present Vera, an end-to-end automated safety testing framework that instantiates software engineering testing principles for non-deterministic agents through a three-stage, self-reinforcing pipeline. First, a literature-driven exploration continuously discovers and structures emerging risks into taxonomies of safety risks, attack methods, and tool execution environments. Second, combinatorial composition across taxonomy dimensions produces executable safety cases, each specifying a concrete safety goal, a programmatically constructed initial state, and a deterministic verification predicate grounded in observable artifacts. Third, adaptive execution runs heterogeneous agents in isolated sandboxes where a control agent steers multi-turn interaction based on runtime observations, while evidence-grounded verifiers judge outcomes from environment state and tool-call evidence rather than model self-report. We evaluate Vera on four production agent frameworks (OpenClaw, Hermes, Codex, Claude Code), revealing substantial safety weaknesses, with average attack success rates reaching 93.9% under multi-channel attacks; we also release Vera-Bench, comprising 1600 executable safety cases spanning 124 risk categories across three execution settings. These results indicate that modular, executable testing infrastructure is essential for rigorous and maintainable safety evaluation of rapidly evolving agentic systems at scale. The code is publicly available at https://github.com/Yunhao-Feng/Vera.

Given one or more images of a railway crossing, can we leverage visual cues that allow us to robustly estimate how safe it is? Can we improve our ability to do so by introducing structured data (such as official accident reports) about the accident history of that crossing into our models? In this work, we explore how to best answer those questions towards building an AI system that can ingest multi-modal data for railway crossings and provide safety assessment and scores that align with expert opinion and with safety scoring used by the Federal Railroad Administration (FRA). To that end, we propose a proof-of-concept pipeline that delivers on that goal, while at the same time exploring and tackling a number of critical research challenges that pertain to different parts of the pipeline, from data preparation to different learning paradigms that can allow us to realize such a system. Indicatively, our proposed system identifies HIGH-RISK and LOW-RISK crossings with a macro F1 score of 0.757 and estimates FRA-based safety scores with an RMSE of 0.078 and correlation of 0.492 using a routed fine-tuned compact VLM pipeline, while producing qualitative results that align with domain-expert assessment.

Current molecular generation benchmarks emphasize task complexity, molecule novelty, and property alignment; they largely overlook a critical concern: the potential safety risks of AI-generated molecules. In practice, many generative models may produce molecules with toxic, reactive, or otherwise hazardous characteristics - posing hidden dangers that remain insufficiently addressed. To address this gap, we introduce MolSafeEval, a benchmark dedicated to evaluating and analyzing the safety risks of molecular generation. Unlike prior approaches that rely on narrow toxicity predictors, MolSafeEval integrates heterogeneous safety knowledge - ranging from toxicological databases to hazard rules - into a structured molecular safety knowledge graph. This graph serves as a foundation for large language model-based reasoning, enabling systematic detection and explanation of unsafe features in generated compounds. We further categorize molecular generative models into four representative task types - unconditional generation, property optimization, target protein-based design, and text-based generation - and provide standardized datasets and safety evaluation protocols for each. By systematically revealing the safety vulnerabilities of current generative approaches, MolSafeEval offers a new lens for benchmarking molecular models and provides essential guidance toward safer, more trustworthy molecular design.

Vision-language models (VLMs) are now proposed as runtime safety guards for embodied agents in homes and factories. A deployable guard must catch genuinely unsafe situations while avoiding unnecessary intervention on routine but superficially alarming activity, a distinction that binary safety benchmarks obscure. We introduce EgoSafetyBench, an egocentric video benchmark of 1,200 robot-view scenarios annotated at half-second granularity, to evaluate VLMs as streaming guards across two tracks. The situational track (800 scenarios) spans four families, from routine and safe-but-suspicious scenes to obvious and contextual hazards. The visual-channel track (400 scenarios) targets in-scene text-a sign, sticker, or label visible in the scene-that can misrepresent the physical situation, pairing each misleading sign with a truthful version to test both whether a guard flags the text as misleading and whether the text corrupts its physical-safety judgment. Both tracks use contrastive ladders: near-identical scenarios differing only in a single visible deciding cue, so a correct call must hinge on that cue rather than the overall scene type. We evaluate ten open- and closed-source VLMs. We find that while guards reliably recognize videos containing hazards, they often miss specific hazardous moments, particularly contextual hazards. Furthermore, misleading in-scene signs degrade all tested guards: vulnerable models miss up to a third of hazards, while robust models over-intervene on safe content. Matched controls reveal that apparent safety robustness often reflects indiscriminate alarming rather than true physical reasoning.

This paper presents a safety-centered empirical evaluation of uncertainty-aware last-layer adaptation for referable diabetic retinopathy screening using RETFound, a self-supervised vision-transformer retinal foundation model used here as a frozen feature encoder, and the public APTOS 2019 and DDR diabetic retinopathy fundus image datasets. We compare a cached-feature softmax head, post-hoc temperature scaling, variational Bayesian last-layer heads, a diagonal Laplace last-layer approximation, and an SNGP-style cached-feature head. On APTOS, uncertainty-aware operating points improved sensitivity and selective-referral behavior. The strongest APTOS selective-referral result deferred approximately 20 percent of cases and reduced accepted-case false negatives to zero while preserving high accepted-case specificity. However, threshold tuning also reduced false negatives at high false-positive cost, so false-negative reduction alone was not unique to Bayesian modeling. On DDR, native Bayesian heads qualitatively reproduced the APTOS direction but with weaker tradeoffs, while the APTOS-trained SNGP checkpoint transferred poorly and failed to provide useful external selective-referral behavior. These results highlight the value of safety-centered evaluation beyond aggregate accuracy: uncertainty-aware last-layer heads can improve internal safety-oriented operating points, but trustworthy retinal screening claims require explicit safety-coverage evaluation and second-dataset validation under shift.

Safety benchmarks often buy scalability by fixing the prompt, the language, and the turn structure. For emotional-support chatbots, that bargain hides precisely where safety failures emerge: across a multilingual, multi-turn crisis conversation. We present EMPATH, a benchmark for safety evaluation of emotional-support chatbots. An auditor model role-plays help-seeking users, generating multi-turn conversations from 140 seed instructions and 34 personas. A judge model scores each full transcript against 19 metrics across five dimensions: crisis handling, therapeutic quality, conversational integrity, emotional safety, and cultural adaptation. EMPATH is built for Mexican Spanish and US English; the studies reported here run in Mexican Spanish. Auditor and judge are drawn from different model families, and the judge is treated as an instrument to be calibrated rather than trusted. A strict per-criterion rubric reveals material score inflation on 10 of the 19 metrics and restores discrimination. We study the measurement properties of the benchmark through judge calibration and cross-family inter-judge agreement. We also illustrate EMPATH on three frontier models, one of them open-weight. Aggregate scores sit within 0.74 points of one another, but per-metric profiles diverge by up to six points in model-specific places. Under the standard rubric, both the ranking and the weak spots are stable across a second, cross-family judge: 93% of scores fall within plus or minus 1. A five-run test-retest adds a second axis: even the steadiest model swings from 2 to 10 on a crisis metric across identical re-runs, and deepseek-v4-pro returns a different conversation on every run even at temperature 0. Run-to-run reliability is therefore a per-model safety property, not noise to average away. EMPATH is system-agnostic; the pipeline, seeds, personas, and rubrics are released for reuse.

LLM evaluation and AI safety face a shared measurement problem: benchmark scores, reward-model signals, and reported safety metrics can improve while the latent properties they are meant to represent remain difficult to verify. This paper combines a hybrid survey - a systematic search paired with narrative synthesis and separately tracked grey evidence - with a conceptual framework and a structured ten-model audit. The synthesis spans eight evidence streams: benchmark validity, dynamic evaluation, LLM-as-judge reliability, safety evaluation, jailbreak/refusal robustness, reward hacking, mechanistic interpretability, and governance/auditability, covering 2018-2026 evaluation-safety measurement work. We introduce EvalSafetyGap as an organizing hypothesis for comparing evaluation-side and alignment-side proxy failures under optimization pressure, using Goodhart's Law together with two constructs we develop here - an Instability Decomposition and an Alignment Trilemma - as tools for generating testable comparisons. The audit shows how conclusions shift when capability, behavioral safety, and governance are measured separately. In this sample ($n = 10$), the association between capability and sustained adversarial robustness is statistically indeterminate using the displayed Table 3 inputs (Pearson $r = +0.232$, $p = 0.520$), and the apparent open-closed safety gap is modest, driven mainly by governance and disclosure rather than behavioral robustness, and sensitive to how a single borderline model is classified; attempt-budget results are protocol dependent. Because the public evidence uses heterogeneous protocols, the audit is diagnostic rather than rank-generating. The contribution is a shared vocabulary and evidence map to support dynamic evaluation, transparent source reporting, multi-attempt safety measurement, and auditable alignment practice.

In embodied intelligence, safety is a prerequisite for reliable robot deployment in the physical world. Current vision-language-action (VLA) models continue to advance toward general-purpose task capability, yet their embodied safety limits remain poorly understood. To address this gap, we introduce ForesightSafety-VLA, a diagnostic benchmark that makes safety the primary evaluation target for VLA systems. We define a 13-category safety taxonomy covering physical interaction safety (Safe-Core), instruction-side safety (Safe-Lang), and perception-side safety (Safe-Vis), and evaluate policies under three controlled dimensions of variation -- scene structure, language command, and visual observation -- so that failure sources can be diagnosed rather than hidden in a single aggregate score. Beyond binary task success, ForesightSafety-VLA measures process-level risk through cumulative safety cost (CC) and risk exposure time (RET), together with a four-quadrant decomposition of safe/unsafe success and failure. We instantiate 66 safety-augmented base scenarios in RoboTwin across 5 embodiments and report results on representative VLA baselines. Across the evaluated baselines, even the strongest policy incurs non-trivial safety cost and unsafe nominal success, while structure and visual variation induce substantially stronger safety degradation than ordinary language variation. These results suggest that embodied safety is tightly coupled to perception, grounding, and control competence rather than being reducible to post-hoc safety filtering alone.

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.

With the widespread adoption of large language models (LLMs) in chatbots and everyday applications, companies increasingly need guardrails that are effective while remaining low-cost and low-latency. Safety evaluation of LLM outputs has generally relied on LLM-based judges, which can be effective but are often slow and expensive to deploy at scale. In this paper, we evaluate whether fine-tuned modern encoder classifiers from the ModernBERT family, including ModernBERT and Ettin, can reliably identify harmful LLM outputs in user-model conversations without substantial performance loss relative to LLM-based judges. We benchmark these encoder classifiers against rule-based prefix matching, fine-tuned LLM classifiers, and LLM judges using a range of judge-prompting strategies across open-source adversarial datasets. The LLM judges include evaluation methodologies from StrongReject, ShieldGemma, JailbreakBench, AILuminate, SorryBench, and a Claude-as-a-judge setup, as well as fine-tuned safety classifiers such as LlamaGuard 3 and LlamaGuard 4. The encoder classifiers are fine-tuned on judge-labeled data using a majority-voting label strategy and are then evaluated on a gold-standard holdout dataset to assess their performance relative to LLM judges. We report absolute performance using F1 score, false negative rate, and precision-recall metrics. We also break down results by attack technique, including single-turn prompting, decomposition, escalation, and context manipulation, to identify where encoder classifiers align with or diverge from LLM-based judges. Our findings provide guidance on when encoder classifiers can serve as cost- and latency-efficient alternatives to LLM-based safety evaluation.

Despite the impressive manipulation capabilities of Vision-Language-Action (VLA) models, their operational safety under strict constraints remains largely unverified. To address this, we introduce a parametric safety benchmark to procedurally generate safety-critical scenarios with comprehensive stochasticity. To overcome the scalability bottlenecks of human teleoperation, we develop a novel keypose-driven data generation pipeline. Leveraging this infrastructure, we curate a large-scale dataset of 19,664 strictly collision-free demonstrations with extensive domain randomization. We then conduct a systematic cross-paradigm evaluation of eight VLA and two embodied foundation models. Our analysis reveals a critical generalization-safety tension: although high-diversity training fosters safer trajectories, task success remains fundamentally bottlenecked by sub-optimal trajectory synthesis and semantic misalignment. By providing a scalable pipeline, a robust dataset, and profound failure-mode insights, LIBERO-Safety establishes a crucial foundation for developing safe and reliable VLA models.

Safety benchmarks assume that test-condition behavior predicts deployment behavior, an assumption that fails if models detect evaluation cues and adapt. This opens a gap between benchmark performance and deployment behavior: compliance measured under test conditions becomes an optimistic upper bound that overstates how safely a model behaves once the evaluation harness is removed. We characterize this evaluation awareness through eight experiments across 37 open-weight models and seven families. (i)Detection is moderate and training-driven (24/37 models exceed chance, best AUROC 0.714 vs.0.819 human, with instruction tuning dominating over scale). (ii)Detection shifts safety behavior (hard refusal drops 5.8 percentage points under hypothetical framing, and 21/140 HarmBench framing effects are significant, with compliance rising up to +30 percentage points. (iii)Representations survive behavioral collapse (probes retain AUROC 0.98 under rewrites that drive behavior below chance, and multi-layer steering causally moves three downstream tasks while random controls do not). (iv)These axes are weakly coupled (only 1/15 correlations are significant, the sole robust link being behavioral detection versus framing resistance, $ρ=-0.79$, $p<0.001$). We call this gap the benchmark illusion: because detectability, behavioral manifestation, and controllability vary independently, it is multivariate rather than a single number, so no single awareness score is a reliable proxy for deployment safety.

Existing driver intervention systems rely on auditory alerts and fixed templates, failing to leverage multi-task recognition outputs. General-purpose metrics such as BLEU and BERTScore cannot capture intervention-specific quality dimensions including risk-urgency alignment, cognitive load, and driver acceptability. In this paper, we propose the Driver Safety-Aware Intervention Score (DSAIS), a domain-specific metric evaluating five dimensions through a hybrid architecture combining lightweight rule-based computation with LLM Judge evaluation, together with an end-to-end framework integrating four-task recognition outputs into an LLM through risk fusion, state history management, and dynamic prompt construction. Experiments on the AIDE dataset with five models and seven conditions demonstrate that DSAIS achieves ICC 0.798-0.840 across three architecturally distinct judges and Cohen's d > 1.5 across all control conditions. Multi-dimensional sub-score analysis quantifies the contextual adaptability gap between rule-based and LLM-based systems, revealing that multi-task integration improves contextual relevance by 9.1% over rule-based baselines. Ablation experiments demonstrate that each framework component contributes to contextual relevance, with sub-score decomposition revealing gains that aggregate scoring masks. Driver emotion recognition is identified as the most critical upstream factor, and compact local LLMs (7B--9B parameters) achieve quality superior to API-based models, providing practical design guidelines for in-vehicle deployment.

Large language model (LLM) agents are increasingly proposed as supervisory components for safety-critical systems, yet their robustness under sustained, adaptive adversarial pressure remains poorly characterized. We present NRT-Bench, a benchmark for multi-turn red-teaming of LLM agents acting as operators of a safety-critical system, instantiated in a simulated nuclear power plant control room. A five-role operator team, each backed by a configurable LLM, runs a plant governed by six critical safety functions (CSFs), while adversaries inject messages over four channels in bounded multi-turn sessions with per-turn feedback. Harm is an objective signal rather than LLM-judged text: a run terminates the moment any CSF is lost, attributed to the causing message. Evaluating four frontier operator models under a fixed-attack paired-replay protocol, we find that adaptive multi-turn attacks reliably push the operator team past a safety limit: across the four models, between 8.7% and 12.1% of attack sessions end with the plant losing a critical safety function. Although the four models look almost equally robust by this aggregate rate, their failures barely overlap: of $149$ sessions, none defeat all four models while a third defeat at least one, so vulnerabilities are nearly disjoint across models rather than nested. The effect of added defences is strongly model-dependent: the same guardrail stack or safety-advisor agent that lowers attack success for one model can raise it for another. We release the simulation venue, attack dataset, and replay tooling for reproducible safety evaluation of LLM agents.

Recent incidents involving LLMs used for mental-health support reveal a critical evaluation gap: surface-level safety scores do not capture how models behave across realistic, emotionally sensitive interactions over time. Existing benchmarks measure knowledge, safety, or static response quality, but miss whether LLM interactions help users keep reflecting, coping, and making decisions themselves. We formalize this missing dimension as COGNITIVE ATROPHY, a process-level behavioural measure in AI-mediated mental-health support distinct from safety and helpfulness. To measure it, we introduce COGNITIVE ATROPHY BENCH, a clinically grounded benchmark built from 1,576 fully human-generated counseling conversations, 15,680 turns, and 42,230 responses from five LLMs. Three clinical and neuropsychology experts developed a 20-attribute schema spanning user context, response behaviour, and global risk flags; six trained clinical reviewers applied it with span-grounded evidence, producing 5,324 reviewer judgments. We further introduce the User-Input Risk Index (UIRI), the Cognitive Atrophy Risk Index (ARI), and trajectory summaries. Across five LLMs, models show a consistent moderate-to-high level of atrophy-aligned behaviour across single and multi-turn settings. While models generally respond to overt safety cues, they adapt less reliably when users seek solutions or decisions. The dominant recurring patterns are directive advice, problem-solving, recommendation responses, topic shifts, and forms of validation that may reinforce dependence rather than reflection. Our work makes COGNITIVE ATROPHY measurable and provides a foundation for auditing model behaviour in sensitive LLM conversations.

Long-tail scenarios remain a major bottleneck for autonomous driving evaluation, even as datasets grow by orders of magnitude. Existing evaluation pipelines are rarely human-aligned, safety-aware, verifiable, and explainable at the same time: closed-loop metrics often saturate among strong planners, while unstructured human ratings can be noisy without a carefully designed protocol. We formulate planning evaluation as additional-threat detection: given a planner trajectory and an expert reference, does the planner's displacement introduce new unsafe driving behavior? We propose FluidTest, an evaluation pipeline with three components: a pairwise WebUI protocol for reliable human annotation; a taxonomy of 32 semantic threats with evidence-grounded decision graphs; and a three-agent verification system with reflection for precision and auditability. Experiments on the WOD-E2E dataset show that FluidTest produces consistent labels among trained annotators and identifies additional threats in 65% of Poutine trajectories and 51% of RAP trajectories. These results show that state-of-the-art planners can still exhibit substantial safety-relevant failures despite high Rater Feedback Scores (RFS) and low Average Displacement Error (ADE). Additional details, guidance, and code are available at https://fluidtest.web.app.

Computer-use agents are increasingly evaluated by whether they complete realistic desktop and web tasks. However, task success alone can miss failures in which an agent reaches the nominal goal through an unsafe shortcut. We introduce OSGuard, a dual-granularity benchmark suite for evaluating safety in computer-use agents under benign, unchanged user instructions. OSGuard contains an action-level benchmark for local guardrail decisions and a risk-augmented execution suite for end-to-end evaluation. The action-level benchmark consists of contextualized proposed actions labeled as allowed, unrelated, or unsafe, each judged relative to the original instruction and current interface state. The execution suite contains manually constructed OSWorld-derived task variants in which the original task remains achievable, but the environment is modified to introduce latent hazards such as destructive overwrites, etc. Each variant is paired with augmented evaluators that retain the original task-success criterion while adding explicit state-based safety invariants, allowing us to distinguish safe completions from unsafe completions that satisfy the nominal task objective. Our experimental results on OSGuard show that current multimodal guardrails can perform well on isolated action judgments, while risk-augmented execution exposes remaining gaps between local oversight and reliable end-to-end safety. This dual-granularity design enables more precise diagnosis of whether models can both recognize unsafe proposed actions and improve full-task safety when deployed as guardrails.

Safety evaluations for dual-use biology assistants often measure base-model capability, refusal behavior, or jailbreak success. These metrics miss a deployment question: for a fixed base model, how does the access condition users actually see change benign utility and harmful actionable assistance? I introduce safeguard-conditioned uplift, a protocol for comparing deployed access conditions through a human-judged utility-risk frontier. I evaluate Claude Sonnet 4.6 and Gemini 3.5 Flash under helpful prompting, safety prompting, and an external safeguarded assistant on a 108-task surrogate benchmark, with the headline claim restricted to a locked 18-task held-out split. In a 600-row blinded human audit, the safeguarded assistant reduces harmful actionability relative to helpful prompting by -0.063 over 49 matched response pairs, with bootstrap 95% interval [-0.117, -0.011], while correctness changes by +0.009 with interval [-0.057, +0.077]. Adaptive, Test-B, cue-ablation, and controller-baseline checks support the measurement story but also show non-dominance: safety prompting is often strongest for Claude, while external control helps more for Gemini and can reduce benign utility. The contribution is not a universal defense. It is a deployment-level evaluation target, plus a learned risk-budgeted calibration procedure, for measuring how user-facing access conditions move the utility-risk frontier.

Quantized checkpoints are often screened first with quality metrics and only later, if at all, with direct safety tests. This paper audits that shortcut on a matched 51-row matrix spanning 6 models, 4 families, a 7-level GGUF ladder, and AWQ/GPTQ INT4 checkpoints. In this matrix the shortcut fails: all 36 quality-safety pairings split direction across models, and 9 hidden-danger rows plus 1 near-hidden-danger row show quality stable or improved while refusal falls by 12-68 percentage points. Seven of the 11 AWQ/GPTQ rows are hidden-danger. A four-probe mechanistic follow-up over the 17 Hugging Face-backed FP16/AWQ/GPTQ cells does not rescue it: entropy, refusal-direction, and calibration probes are weak or null separators of dangerous rows, and although probe-identified safety-associated neurons absorb 1.39$\times$ more quantization error overall ($p < 5 \times 10^{-7}$), the effect is not regime-specific. Claude Sonnet 4 relabels 11,470 items in a predefined stratified set, agrees with the primary gemma3:12b judge on 89.9% of rows ($κ= 0.873$, 95% CI [0.866, 0.881]), and changes 0/10 hidden-danger cells. A calibrated study-internal behavioral screen -- the Refusal Template Stability Index (RTSI), built from four refusal-template drift features and calibrated on this matrix -- routes 10/10 hidden- or near-hidden-danger rows to direct safety testing (Wilson 95% CI lower bound 0.72) while leaving 23 of 45 non-baseline rows in a low-risk bucket under both in-sample scoring and row-level leave-one-out validation; on the same matrix, the best single-feature baselines (unique-prefix-rate-delta, raw refusal-rate delta) recover 9/10 and 8/10 respectively at matched bucket size, and cross-stack transfer requires recalibration. For the quantized checkpoints, model families, and safety outcomes studied here, retained quality cannot waive direct safety evaluation.

Safety judges are increasingly deployed to evaluate model outputs against evolving criteria, yet recent meta-evaluation work shows they remain brittle under prompt and rubric variation, with false negative-rate swings of up to 0.24 reported for stylistic perturbations alone. We argue that safety judgment is fundamentally a rubric-following problem: a robust judge must apply the given evaluation criteria consistently across rubric formulations rather than memorize one specific template. We propose a training strategy that combines (i) instance-conditioned dynamic rubrics generated from prompt-response-label triples to expose the judge to the variability of evaluation criteria, and (ii) a reliable-to-expressive curriculum that begins with clean fixed-rubric supervision and progressively introduces noisier dynamic-rubric data. We evaluate on a single human-labeled set under three contrasting rubric prompts (HarmBench-style, ShieldGemma-style, and a domain-specific rubric). Our 12B curriculum judge achieves 94.12-94.88% accuracy across the three rubrics with a cross-rubric range of only 0.76, outperforming general-purpose LLMs, dedicated safety classifiers, and reasoning-oriented judges up to 30B in both peak accuracy and stability. An ablation shows that naively mixing dynamic rubrics into SFT increases cross rubric variance (1.44 -> 3.60); only the curriculum schedule recovers and improves on the fixed rubric baseline (variance 0.76).

Large language models (LLMs) are increasingly evolving from simple text-based interaction systems into LLM agents that can maintain memory, use tools, access external environments, and execute tasks. As their capabilities and autonomy expand, the safety risks they face also become more diverse. Existing evaluations often rely on manually written scenarios, static prompts, or final-output judgments, making it difficult to capture the diverse risks that agents may face during task execution. We introduce ForesightSafety-SAGE, a fully automated scenario generation and safety evaluation framework for LLM agents. Based on five risk dimensions,we instantiae abstract and diverse safety risks in real-world task execution into 1,072 measurable evaluation scenarios. Using the automated evaluation pipeline, 12 LLM agents are evaluated under two authority contexts. The results show that current agents still face substantial behavioral safety risks during task execution, with an average ASR of 47.1% and several models exceeding 70%. These findings demonstrate the importance of executable, process-level evaluation for understanding and improving LLM agent safety.

Are tool-calling LLM agents equally safe throughout a conversation? We discover they are not: agents are most vulnerable at the very start of a session and become substantially safer after a few regular agentic tasks -- a phenomenon we term the cold-start safety gap. To study this systematically, we introduce Safety Over Depth for Agents (SODA), a benchmark that controls how many regular agentic tasks the agent completes before encountering a safety threat, supporting up to 20 preceding tasks. Evaluating 7 models from 4 families, safety improves by 9--52% as the number of preceding regular agentic tasks increases from zero to twenty. Representation analysis confirms that model hidden states gradually shift toward a safety-aligned region as more preceding tasks are present. By systematically studying which part of the preceding conversation matters most, we find that the regular agentic tasks themselves are the primary driver of safety, while the agent's own prior responses have less effect on safety but are essential for preserving later utility. This conclusion is further supported by evaluation on open-source safety benchmarks (AgentHarm, Agent Safety Bench) and utility benchmarks (BFCL, API-Bank), confirming that warming up the agent with regular agentic tasks before deployment makes it safer and preserves full capability. Based on these findings, we recommend a simple deployment strategy: having the agent complete a few regular agentic tasks before possible exposure to safety-critical requests mitigates the cold-start safety gap. Our code is available at https://github.com/Trustworthy-ML-Lab/Agent-Cold-Start-Safety-Gap

Embodied AI systems are increasingly expected to reason and act over extended horizons in physical environments. This growing capability brings safety to the foreground, because failures in the physical world can harm people, damage objects, and disrupt workplaces. Although safe embodied AI has attracted substantial attention, the literature remains fragmented across planning, policy design, and runtime execution. Long-horizon robotic manipulation is a particularly revealing anchor domain for this problem because semantic misgrounding, subtask-level error propagation, execution drift, and contact-rich physical risk can accumulate within the same closed-loop system. This survey therefore provides a structured review of safety in long-horizon robotic manipulation from an embodied AI perspective. We organize the literature by intervention locus, covering planning-time, policy-time, and execution-time safety, and we analyze the strength of the evidence that each line of work provides, distinguishing formal guarantees, statistical support, and empirical safety heuristics. This framework clarifies the distinct roles of backbone capability papers, direct safety mechanisms, and benchmark or evaluation studies, while exposing where current safety claims are well supported and where they remain indirect. We identify persistent gaps, including limited evidence for policy-time safety, weak formal support for contact-rich long-horizon manipulation, immature uncertainty-triggered intervention, and a shortage of manipulation-specific safety benchmarks. We conclude by outlining research directions for cross-layer assurance, evaluation design, and safer deployment of long-horizon robotic agents in real-world settings.

Patient safety event triage, determining whether a clinical event is reportable under jurisdiction-specific policy, is a high-stakes task typically performed manually by patient safety experts. Although LLMs may support this workflow, reliable evaluation is limited by the lack of benchmarks to capture evidence-grounded policy reasoning, proactive information seeking for incomplete reports, and principled abstention in irreducibly ambiguous cases. We address this gap with a policy-grounded construction methodology centered on the clause card, a structured representation that factorizes regulatory text into auditable decision specifications. Combining clause cards with anchor-driven instantiation and closed-loop verification, our scalable pipeline produces narratives with by-construction ground truth and naturally supports generating missing information and uncertain variants. We instantiate this method on Minnesota's 29 Reportable Adverse Health Events, producing PSEBench, a 5,074-case benchmark with an agentic evaluation environment. Evaluation on 15 representative LLMs reveals consistent capability trends, demonstrates the benchmark's utility, and identifies actionable gaps toward reliable LLM-based patient safety event triage.

An attacker that strategically chooses when to attack is much harder to catch than one that attacks indiscriminately. AI control is a safety framework for deploying capable but untrusted AI agents under the oversight of a weaker, trusted monitor and a limited human audit budget. Control evaluations stress-test these protocols by pitting a red-team attack policy against the blue-team monitor, but current evaluations typically assume attackers that do not strategically select when to attack. We study this capability, attack selection, in agentic settings by decomposing attack decisions into a start policy, which decides when an attacker should attack, and a stop policy, which decides when an attacker should abort an ongoing attack. Across two agentic settings, BashArena and LinuxArena, both policies substantially lower measured empirical safety without changing the underlying attack capability. At a 1% audit budget, our start policy reduces safety by 20pp on both BashArena and LinuxArena, and our stop policy reduces safety by 20pp on BashArena and 28pp on LinuxArena. These reductions should be interpreted as upper bounds on the effect of attack selection. Existing control evaluations may therefore yield overly optimistic safety estimates against selective attackers. We recommend that future evaluations, system cards, and safety cases elicit attack selection to produce more realistic safety estimates.

Adapting foundation large language models to a user's task or preferred style through fine-tuning can result in compromising the model's safety. Previous works examined the effects of fine-tuning on model safety in limited and seemingly random experimental settings. We argue that anchoring fine-tuning to a specific capability goal is essential for avoiding arbitrary empirical choices, allowing us to draw meaningful conclusions about safety impacts, and to compare mitigation methods on a consistent basis. We conduct a multi-dimensional evaluation of the effects of fine-tuning on model behavior by focusing on capability as well as safety. Our results surface important issues that (1) fine-tuned models can produce incoherent generations in response to safety prompts, (2) automated safety judgments are unreliable for such incoherent outputs, and (3) the conclusions about the effects of fine-tuning can change depending on the choice of safety benchmark as well as the safety evaluator.

Benchmarks for autonomous agents measure whether agents complete tasks, yet this framing is systematically blind to whether an agent should have proceeded at all. Agents trained under human-feedback objectives develop a structural tendency to proceed even when they lack the inputs, evidence, or authorization to act safely, a disposition we term compliance bias, because both the reward signal and the benchmark scoring regime treat proceeding as the correct default regardless of whether the preconditions for safe action are present. We make three contributions. We first show that compliance bias originates in reward hacking within human-feedback pipelines and is entrenched by prominent agent benchmarks, which either penalize agents for pausing or are architecturally unable to distinguish a principled pause from a silent failure. We then introduce a three-gap taxonomy of abstention-warranted scenarios, covering specification gaps where required information is absent, verification gaps where world state cannot be confirmed, and authority gaps where explicit authorization has not been given, which together provide a principled basis for constructing abstention-aware agent benchmarks. Finally, we propose abstention evaluation protocols (Safety Rate, Usability Rate, and Informed Refusal Rate) and report preliminary results across 144 enterprise agent scenarios and five model families, in which a runtime-enforced abstention mechanism achieves up to 89.2% hazardous-action blocking and 87.5% usability on authorized scenarios, demonstrating that the safety--usability tradeoff is tunable rather than inherent and that its shape varies substantially across model families. We treat this as preliminary work and offer the taxonomy and composite metrics as a starting point for further conversations.

Safe global deployment of AI models requires alignment with human values that vary across cultures. Yet rater pools in safety evaluation datasets remain largely geographically homogeneous, failing to capture geo-cultural differences. Further, it remains unclear whether such differences persist after controlling for demographics such as age, gender, and ethnicity. Through a meta-analysis of safety datasets, we find that most do not report geo-cultural information, and those that do lack a unified methodology to jointly analyze geo-cultural and demographic correlates. Using the Inglehart-Welzel dimensions of cross-cultural variation, we demonstrate via multilevel modeling that cultural zone membership explains variance in safety ratings beyond standard demographics (p<0.05 across 6 datasets). Moreover, our analysis indicates that roughly 10% of items in the datasets we examined are culturally sensitive: likely to be misclassified as safe without adequate cultural representation. We evaluate LLMs as both rater surrogates and triage tools, finding that current LLMs do not reliably stand in for raters, though they can help prioritize culturally sensitive items for human annotation. Our findings motivate more culturally pluralistic safety evaluation and offer practical takeaways to support it.

We evaluate the consistency of automated judges in conducting a multi-dimensional safety evaluation in a reference-free setup. Our results indicate that Large Language Models are unreliable judges in identifying safety issues related to machine-generated advice in regulated domains such as finance, although they are more reliable at identifying more overt forms of unsafe/harmful content such as violence. The degree of inconsistency in a model's judgments can vary significantly by the chosen safety criteria and can be impacted by the language of the content and its linguistic style as well. Finally, there is high disagreement among different judges for the same output, across domains, safety criteria, and languages. These findings provide new insights on the practice of using LLMs as evaluators and offer several recommendations for practitioners on how to use automated judges in practical scenarios.

Aligning large language models (LLMs) to heterogeneous and rapidly evolving safety requirements remains a critical challenge. Existing instruction-tuned LLMs and standalone safety classifiers often fail to generalize to new safety configurations, motivating the need for Reward Models (RMs) that are explicitly configurable to changing specifications. We introduce the Configurable Safety Reward Model (CSRM), which is jointly optimized for calibrated safety compliance and reward modeling. Our approach is supported by configuration-targeted data augmentation that enforces instruction adherence while preserving relative severity structure. The resulting RM is sensitive to fine-grained safety configurations and conversational nuances, substantially improving generalization to previously unseen safety configurations. CSRM achieves state-of-the-art performance on recent configurable safety benchmarks, including CoSApien (94.6% F1) and DynaBench (75.8% F1), without requiring additional human annotation. When used for downstream safety alignment, CSRM yields LLMs with a significantly improved helpfulness-safety tradeoff compared to existing baselines.

When Large Language Models (LLMs) are deployed in Chinese-language settings, a troubling pattern emerges: safety systems that work well in English break down. These systems struggle to cross linguistic and cultural bound-aries, leaving models exposed to adversarial prompts that exploit Chinese-specific evasion techniques, including Pinyin romanization, character decomposition, internet slang, and hedging tone. To address this gap, we introduce ChiSafe-PAS (Chinese Safety Pilot Annotation Set), a human-annotated benchmark of 1,897 adversarial Chinese prompts spanning four high-stakes domains: self-harm and violence, drug and illicit trade, fraud, and satire. Of these, 1,544 entries carry complete gold-standard annotations: a 3-class response label (REFUSE, SAFE-REDIRECT, RESPOND), a nine-category obfuscation taxonomy, a risk-level rating, and annotator rationale. We describe the dataset design, annotation process, and obfuscation taxonomy in detail. Our primary goal is practical: to give the research community a high-quality, culturally grounded resource for benchmarking LLM safety alignment. In doing so, we engage three broader tensions in the field: the blurring boundary between training and evaluation data, the need for domain coverage grounded in real-world risk, and the limits of scale as a substitute for cultural expertise.

Vision-Language-Action (VLA) models with integrated reasoning have been proposed for end-to-end autonomous driving, assuming a tight coupling between reasoning and trajectory generation. However, the robustness of such systems under realistic input perturbations remains largely unexplored. We show that these models are highly vulnerable to realistic input perturbations, achieving up to 89% attack success rate (ASR) on reasoning and up to 72% on trajectory manipulation in closed-loop simulation, leading to increased collision rates and degraded safety metrics. Using NVIDIA's recent Alpamayo models as representative industry-developed VLAs, we conduct the first systematic black-box study of reasoning-enabled VLA models under realistic textual input corruptions, evaluating their impact on reasoning and driving behavior. We introduce a reasoning-aware evaluation framework capturing both semantic and structural aspects of reasoning, along with safety-centric measures. We also introduce a benchmark for evaluating attacks and defenses on reasoning-trajectory interactions in autonomous driving. Our results highlight the need for rigorous evaluation and improved defenses to ensure the safety of reasoning-enabled VLA systems in autonomous driving.

The validity of AI safety evaluations depends on models behaving consistently across controlled and deployment settings. Prior work has identified test-time contextual cues, such as hypothetical scenarios, as a source of verbalized evaluation awareness and subsequent behavioral shift. In this paper, we investigate a potential explanation of this phenomenon: evaluation meta-knowledge, defined as parametric knowledge about the structural traits that characterize evaluations. Similar to dataset contamination, where benchmark exposure leads to higher performance through memorization, we hypothesize that models trained on texts describing evaluation practices may implicitly learn to recognize and respond to evaluation-like contexts, for instance, through exposure to scientific articles or social media posts about AI benchmarking. To test this, we fine-tune models on synthetic documents describing evaluation traits such as verifiable structures or harmful requests. Evaluating this fine-tuned model on five safety benchmarks, we find that it is significantly safer than the base model and control model. This behavioral shift persists even when restricting the analysis to responses lacking explicit verbalization of evaluation awareness. Our results demonstrate that evaluation meta-knowledge may inflate safety benchmark performance, introducing a novel confounder that is independent of explicit memorization or verbalized evaluation awareness, thus, challenging to detect. These findings have important implications for the design and interpretation of AI safety evaluations. Our code and models are available at https://github.com/compass-group-tue/arxiv2026_evaluation_meta_knowledge.

Safety benchmarks for language models are typically evaluated using static paradigms that treat all items as equally informative for all models, an assumption that is particularly problematic for adversarial, highly heterogeneous safety items. Applied in full to modern benchmark suites, the current evaluation procedures would require on the order of $10^5$ responses, most of which provide little ranking signal. We analyze a suite of widely used safety benchmarks and make three contributions toward more efficient safety evaluation. First, we show that Item Response Theory (IRT) recovers interpretable structure on safety benchmarks, with ability estimates resolving differences among models that cluster at the ceiling of raw safety metrics. Second, we show that adaptive item selection, which dynamically chooses informative items for each model based on its responses, approximates full-benchmark rankings while reducing evaluation cost by at least 80% on benchmarks where Spearman's $ρ>$90% with full-benchmark is attainable, and by up to 99.9% on AIR-Bench 2024. Third, we introduce a practical procedure for extracting a fixed, informative subset of items reusable across models, providing an alternative to adaptive selection with savings of up to 99.8% on AIR-Bench 2024. Together, these results establish that psychometric methods enable benchmark-aware reductions in evaluation costs across the safety evaluation pipeline.

Testing object detectors in safety-critical domains requires semantically meaningful probes beyond pixel-level corruptions. We present SemProbe, a tool for semantic robustness probing: users upload deployment images, create masks manually or automatically, select operational design domain-derived factors (or custom prompts), and run diffusion-based controlled inpainting. The system supports batch jobs, parallel seed/workflow variations, and configurable generation parameters. After each output, model inference runs automatically and displays annotated before/after comparisons with performance deltas. All probes are logged as structured artifacts, enabling traceable robustness evidence aligned with safety evaluation workflows. We demonstrate \textsc{SemProbe} on hand detection for dimension saws, targeting factors from insurance-oriented test criteria.

Kazakh is underrepresented in resources for evaluating the safety behavior of large language models. We present KZ-SafetyPrompts, a Kazakh prompt dataset for safety evaluation across eleven categories covering common risk areas such as self-harm, violence, child exploitation, sexual content, racist content, radicalization, and regulated goods or illegal activities. The dataset contains 5,717 prompts written natively in Kazakh (Cyrillic), organized by category, with English translations for cross-lingual analysis. Prompts resemble realistic user queries, often in a teen or child style, and are phrased as intent prompts without procedural instructions. We document the writing protocol, labeling procedures (including borderline-case decision rules), and quality-control steps (schema standardization, completeness checks, and deduplication). We also align the categories with widely used safety taxonomies to support integration with existing evaluation pipelines. Baseline results with GPT-4o show an overall refusal rate of 28.2%, varying from 5.5% to 53.8% across categories, indicating that Kazakh prompts expose category-specific safety gaps not captured by English-only evaluation.