Before the Script, Set the Stage: How Worldview Simulation Amplifies Psychologically Grounded Persuasion in Multi-Turn Jailbreaking

Multi-turn jailbreak attacks demonstrate that harmful intent can be distributed across dialogue, yet existing methods obscure what conversational mechanisms drive vulnerability. We introduce BLUEPRINT, a safety-evaluation framework separating a factorized social-influence strategy space from WORLDVIEWSIM, a cross-turn situational context module. Monte Carlo Tree Search optimizes turn-level combinations of 18 theory-grounded influence factors across a four-turn trajectory. Across six frontier models, BLUEPRINT achieves near-ceiling ASR on major open-weight and proprietary models, while requiring the fewest average queries (2.46). The resulting trajectories further reveal model-specific vulnerability among resistant targets: each responds to distinct influence factors and strategy transitions, yet all share a common recovery pathway-shifting toward concrete, executable task framing consistently escapes hard-refusal states. Ablations confirm operational cues matter most: making requests actionable has the largest impact, gain framing is unusually potent, and some legitimacy appeals can backfire. These findings suggest robust multi-turn safety requires monitoring not only harmful content, but also how dialogue state makes unsafe requests appear concrete and locally executable.

FUSE: An Evaluating Framework for Dangerous Capabilities of LLMs

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 $ρ\in [0.32, 0.52]$).

SDARE-Bench: Evaluating Large Language Models on Conversational Stigma Detection and Response in Dyadic and Group Dialogue

Large Language Models (LLMs) are increasingly used in advice seeking and decision making that may affect social judgements. Despite stigma's profound effects on people and communities, benchmarks remain scarce. Existing general-domain evaluations typically rely on static prompts and fixed-format tasks, overlooking conversational contexts and audience effects in everyday communication. To address these gaps, we introduce SDARE-Bench, the first scenario-based benchmark evaluating both stigma detection and open-ended response generation in LLMs, comprising 1,138 dyadic queries and 1,388 group dialogue. Empirical results across 8 LLMs consistently demonstrate poor identification of stigma components, especially in group dialogues. In open-ended response generation, stigma expression was substantially higher in group settings than in dyadic, with weaker resistance to stigma and more unrealistic advice. Responses were evaluated using a classifier trained on 1,392 human annotated responses. In constructed group pressure settings, stigma expression rates further increased to a striking average of 97.5%. Our findings identify stigma response as a recurring LLM safety vulnerability, especially in socially complex conversational contexts.

Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents

Skill-augmented agents load reusable skills as persistent runtime context, improving task performance but also giving malicious skills a durable channel for steering future actions. Such skills may leak secrets, corrupt code, bypass approvals, or stage data for exfiltration only after a concrete user task and workspace state make the unsafe action appear useful. This makes pre-install vetting insufficient and calls for runtime, task-conditioned protection. We propose Defense-as-Skill, a defense paradigm that implements the runtime guard itself as an installable, inspectable, and editable skill. Our guard, SkillSonar, runs alongside untrusted task skills and checks sensitive actions against the user's task boundary, routing each action to an allow, replan, or confirmation decision without modifying the underlying agent runtime. To study this setting, we construct SCOPE-R, a task-conditioned dataset covering 6 risk families and 21 sub-categories, with 206 attack-confirmed malicious instances and 43 benign tasks. We then improve SkillSonar on the SCOPE-R training subset using runtime guard-skill evolution, a Monte-Carlo Tree Search procedure that evolves the on-disk guard skill from feedback on the rollouts. Across Claude Code and OpenClaw, the evolved guard substantially reduces attack success while maintaining a favorable safety-utility trade-off. On repeated GLM-5 runs, SkillSonar reduces ID ASR from 0.482 to 0.104 and OOD ASR from 0.606 to 0.115. Further analyses demonstrate transfer across victim models, held-out risk families, and external benchmarks, as well as retained protection against adaptive attackers. Ablations further show that explicit safety responsibility assignment and the skill-native representation are both important to the observed gains.

Same Request, Different Boundary: Evaluating Cybersecurity Assistance across Conversational Contexts

Large Language Models (LLMs) can solve complex problems, but their misuse in high-risk domains can lead to severe consequences. Model providers therefore restrict assistance for potentially harmful requests. Refusing all cybersecurity requests would therefore harm legitimate users. Providers need a mechanism to block malicious use without denying legitimate assistance to defenders. Existing cybersecurity-specific datasets evaluate this mechanism, but none considers the conversational context of a request. We introduce 3R-Bench (Refusal, Repetition, and Revision), a benchmark of 150 real-world cybersecurity requests augmented with two adversarial conversational settings, and evaluate eight LLMs on it. Prior assistant behavior strongly changes responses to an unchanged request: among 376 available pairs from a 400-pair panel, compliance rises from 62.0% after refused history to 85.1% after accepted history. The opposite pattern appears under dialogue decomposition. In comparison, compliance falls from 501/800 direct responses to 172/800 after dialogue; among 738 pairs returning model-authored text in both conditions, the decrease is 45.1 points. Failure feedback recovers only a small fraction of this loss.

BLOOM-WILT: Logit Tilting for Behaviour Elicitation in Automated LLM Auditing

Users of a deployed language model routinely encounter behaviours that testing almost never surfaces, since deployment puts the model through orders of magnitude more interactions than any evaluation can simulate. Automated auditors make testing cheap to scale and flexible enough to cover almost any specified behaviour, yet their lack of optimisation pressure makes them sample-inefficient. To address this shortcoming, we introduce BLOOM-WILT, a full auditing pipeline that elicits natural multi-turn instances of rare behaviours, without training cost or access beyond the target's next-token distribution. On the input side, WILT's auditor model revises its conversational strategy across rounds, learning from previous scored interactions. On the output side, WILT adaptively reweights the target's decoding using the model's own distribution conditioned on an elicitation prompt, so that behaviour-relevant generations are sampled ahead of others it finds equally probable when unprompted. We evaluate WILT across 4 target models and 8 behaviours, where it beats the baseline auditor in 30 of the 32 settings and overturns the previous model safety rankings. WILT raises average behaviour presence from 51% to 100% when eliciting self-harm encouragement from Qwen3.5-4B, beating every elicitation method we port into the same pipeline at matched compute, without pushing output probability below the baseline's.

Reference-Grafting Matches Fine-Tuning at Eliciting Sandbagged Capabilities

Sandbagging, in which a model deliberately underperforms on an evaluation despite retaining the underlying capability, threatens the safety evaluations that frontier-model governance depends on. The Elicitation Game found that fine-tuning elicits hidden capability from sandbagging model organisms whereas additive activation steering fails. We revisit that verdict with reference-grafting, which sets an activation's coordinate along a contrast direction to the value it takes in an honest reference, at a small set of circuits chosen by active learning. Across eleven fine-tuned password-locked models (1.5-32B, three architecture lineages), it recovers +94 to +101% of the honest-sandbagging gap, matching fine-tuning elicitation without weight updates or training labels; two to five paired examples suffice to fit the direction. Similar recovery holds for reinforcement-learning-induced sandbagging and for password-locked code generation. Grafting works because the fine-tuned lock is a thresholded gate: held-out accuracy stays at the sandbagged level until the grafted coordinate crosses a threshold near the honest reference, which is why additive steering and zeroing the coordinate often fail. The direction tracks how the sandbagging was induced rather than what is withheld -- fit on grade-school science (ARC) it elicits withheld hazardous knowledge (WMDP), yet prompting, fine-tuning, and reinforcement learning each carry a different direction. Circuit-breaking marks the boundary: it reroutes activations on every forward pass, so the fixed edits we test are re-broken downstream and do not restore coherent generation.

Benevolent Bias in Multi-Turn Human-Agent Dialogue

Bias in human-agent interaction can manifest not only through hostile language but also as benevolent bias, whereby unequal treatment hides behind a warm, positive tone. To make it detectable, we operationalise benevolent bias along two dimensions, tone and treatment, yielding three classes: neutral support, overt bias, and benevolent bias. Building on these definitions, we construct BENEVDIAL, a class-balanced corpus of 362,880 multi-turn support dialogues spanning user and agent demographics, roles, and generators, to support controlled evaluation. We then test two detector families on it: off-the-shelf safety detectors and prompted large language model (LLM) judges. Our results reveal a detection gap: off-the-shelf detectors reliably flag overt bias yet largely miss benevolent bias, while LLM judges catch more under more explicit detection criteria but increasingly misclassify neutral support as benevolent bias, and demographic context amplifies the false alarms. These findings suggest that fair monitoring of human-agent dialogue must look beyond surface cues to whether the agent's treatment is disparate.

REINS: Refusal-Enhanced Inhibitory Steering with Sparse Autoencoder Features

Steering with Sparse Autoencoders (SAEs) offers a lightweight inference-time path for adapting the behavior of large language models without retraining. By exposing sparse and interpretable features, SAE steering provides a promising interface for safety control that guides harmful continuations toward refusal. However, we observe that complex wrappers can still undermine existing SAE steering methods on harmful prompts. To evaluate this failure mode systematically, we construct Generalized Undercover Instruction Safety Evaluation (GUISE), a dataset of harmful prompts with complex wrappers. Existing single direction SAE steering methods do not reliably produce refusals on harmful prompts, suggesting that refusal enhancement alone can be too weak when the harmful continuation path remains active. This motivates us to propose Refusal-Enhanced INhibitory Steering (REINS), which suppresses harmful continuation features and enhances safe refusal features in the same SAE feature space. Experiments on GUISE and other datasets show that prior methods either intervene too weakly or achieve only apparent safety through collapse, while REINS substantially reduces harmful responses, markedly improves safe refusals and largely preserves general capabilities.

Evaluating the Safety of Deep Learning-Based Brain MRI Reconstruction

Objective: Deep learning accelerates brain MRI four- to tenfold, but models can erase lesions or synthesize false tissue - failures pixel-averaged metrics like PSNR and SSIM miss. We review whether current evaluation practices detect this blind spot. Methods: Following PRISMA 2020, we searched seven databases without date limits, including 263 studies (1995-2026), appraised them using QUADAS-2 and matched instruments, and synthesized narratively. Categories were derived from titles, abstracts, and controlled vocabulary; reported prevalence figures represent floors. Duplicate screening achieved high agreement (Fleiss kappa = 0.877), as did appraisal (0.788; 0.390 where observable). Extraction is unaudited. Results: Only 18 of 263 studies (6.8%) recorded both a fidelity metric and reader assessment on identical data, leaving the central surrogate unmeasured. Reader studies mostly measured inter-reader agreement, which was weak: fastMRI 2020 concordance reached 0.457 and 0.386 (Kendall W), improving only where SSIM diverged. Erasing a 100 mm3 lacunar infarct shifts global PSNR by 0.03 dB under the stated error model. As the corpus grew fivefold, reader assessments dropped from 32% to 18%, recovering to 21%. Generative models - most associated with hallucination (39%) - were among the least reader-evaluated (11.3%), while self-supervised models reached 47% with zero reader evaluation. Only 5% released code and ran reader studies; none evaluated a model observer; no named dataset covered acute stroke or hemorrhage. Conclusions: On these floors, current evaluation practices cannot certify diagnostic safety. We derive five requirements safety-oriented evaluations must meet.

NeuronFuzz: Safety Neuron Guided Fuzzing for LLM Safety Evaluation

Safety evaluation is critical for assessing whether aligned Large Language Models (LLMs) remain robust against jailbreak attacks. Existing automated testing methods, however, largely rely on response-level feedback: each candidate prompt typically requires generating a target-model response to evaluate its attack effectiveness. This process is expensive and, more importantly, provides only sparse guidance on strongly aligned models, where most candidates are rejected with the same failure outcome. This paper presents NeuronFuzz, a white-box fuzzing framework that exploits internal safety neurons as continuous execution feedback for LLM safety evaluation. A SafetyOracle converts safety-neuron activations into a continuous safety alarm score that serves as feedback for fuzzing and can be obtained during prefill, eliminating response generation from the fuzzing loop. To construct the SafetyOracle, NeuronFuzz uses template-invariant harmful and benign inputs and stability-aware selection to identify a compact set of safety neurons whose activations capture harmful-intent recognition. Moreover, since the safety alarm score is differentiable, NeuronFuzz uses its gradients to identify safety-sensitive template positions and a masked language model to generate fluent, context-compatible mutations while preserving original harmful payload and avoiding additional optimization variables. We evaluate NeuronFuzz across 21 text and multimodal models. Across five white-box source models, it achieves a 76-100% jailbreak discovery rate, outperforming baselines by up to 48 percentage points. Its optimized templates further transfer zero-shot to open-weight and six proprietary target models, achieving average ASR and top-5 ensemble ASR (EASR) of 69.6%/92.6% and 44.1%/60.0%, respectively.

ADeptS-Bench: Measuring the Trustworthiness of Computer Use Agents Across Devices

Computer Use Agents (CUAs) are increasingly deployed to navigate mobile and desktop applications on behalf of users, yet no benchmark comprehensively evaluates whether they can safely interact with visual interfaces while handling ambiguous instructions. We introduce ADeptS-Bench, a dual-stream trustworthiness benchmark, grounded in the ADEPTS capability framework and general population user studies. The Safety stream provides paired benign/malicious tasks with threats embedded in the visual interface. The Disambiguation stream evaluates whether agents seek clarification when intent is ambiguous. Evaluating seven models reveals that no model consistently exceeds 80% task success while staying below 30% attack success; every model clicks "Checkout" on a $25K order without hesitation, and none detects that a "factory reset" button is mislabeled as "Optimize." An ablation reveals three distinct safety architectures: tool-dependent (ASR +21-23pp without refusal tool), partially tool-dependent (+10-11pp), and no mechanism (unchanged). In disambiguation, all models overestimate consequence severity, mirroring the over-refusal bias observed in safety. We release all data, evaluation code, and analysis tools upon publication.

TRACE: An Evidence-Grounded Benchmark for Safety Evaluation of Large Reasoning Models

Large Reasoning Models (LRMs) generate intermediate reasoning traces that may contain unsafe content, even when their final responses appear safe. Guardrail models are designed to detect and block unsafe content, yet existing benchmarks for unsafe content detection focus primarily on prompts and final responses, leaving reasoning traces largely unexamined. Moreover, these benchmarks typically provide only binary safety labels, without evidence annotations that justify the judgments. To address these limitations, we introduce TRACE, an evidence-grounded safety evaluation benchmark that covers the entire LRM inference pipeline: prompts, reasoning traces, and final responses. TRACE includes prompts in two languages spanning nine risk categories and ten attack strategies. For each prompt, four LRMs generate reasoning traces and final responses, and we annotate the safety of each component and extract supporting evidence from the corresponding source text. Evaluating 18 guardrail models on TRACE reveals that safety judgment for reasoning traces is substantially more challenging than for prompts or final responses, and that current models struggle to accurately extract supporting evidence. These findings highlight the need for guardrail models that can reliably detect and precisely localize unsafe content across the LRM inference pipeline.

'Ghaib in Translation' aka Unseen Harm: Measuring Cross-Script Safety Inconsistency with 'Missed-in-Urdu' Scores in LLM Hate Speech Detection

Urdu, the world's tenth most spoken language with 246 million speakers, remains almost entirely absent from mainstream LLM safety evaluation and nine years of WOAH proceedings. To investigate whether this absence has measurable consequences for content moderation reliability, five large language models, GPT-4o, Claude Sonnet 4.5, Gemini 2.5 Flash, Qwen-2.5, and Llama-3.1, were tested across six datasets spanning Nastaliq Urdu, Roman Urdu, English, and code-switched Urdu-English. Across the five Urdu-script datasets, label instability between original-script and English-translation classification ranged from 15.9% (Gemini 2.5 Flash) to 31.6% (Qwen-2.5), with a 'Missed-in-Urdu' rate, content flagged as harmful in English translation but passed as normal in the original script, ranging from 2.4% to 9.9% (median 4.3%). A complete enumeration of all 205 papers across nine ALW/WOAH editions via the ACL Anthology API confirms zero dedicated Urdu papers across the entire period. Results indicate that current LLMs provide uneven safety assurance across Urdu's script varieties, with smaller open-weight models showing substantially higher instability and missed-harm rates than frontier closed models.

EviSafe: Evidence-Grounded Safety Evaluation for Vision-Language Models

Vision-language model safety benchmarks typically evaluate only final responses: whether a model refuses, warns, or complies. This outcome-level view cannot tell whether a model is safe for the right multimodal reason. Safelooking behavior may reflect keyword-triggered refusal, missed visual hazards, or over-refusal of benign-sensitive inputs. We introduce EviSafe, an evidence-grounded framework for VLM safety that jointly evaluates natural user-facing behavior, explicit grounding in textual and visual evidence, and behavioral sensitivity to counterfactual changes in safety-critical evidence. EviSafeBench instantiates the framework as a controlled benchmark with 1,181 gold image-text scenarios and 2,452 targeted counterfactual variants across eight safety domains and eight risk-source types. Each scenario includes a gold safety decision, evidence annotations, a safe-response policy, and counterfactual interventions. The three-probe protocol queries models with natural-response, evidencereporting, and counterfactual-response prompts, then scores them using an evidence-aware judge. Across eleven evaluated VLMs, natural severity accuracy ranges from 27.6% to 52.8%, relaxed diagnostic consistency from 6.1% to 29.3%, and unsafe-to-safe counterfactual transition success from 30.4% to 58.4%. These gaps show that the evaluated VLMs are not reliably safe for the right multimodal reason and motivate evaluation beyond refusal counts.

Stress Testing Unlearning Algorithms

Recently, machine unlearning, the removal of specific training data influence from a model, has gained increasing attention. In large language models (LLMs), unlearning is particularly challenging due to the ambiguity of inputs and outputs. Con- sequently, rigorous evaluation is critical for assessing both safety and utility, and for driving progress in unlearning meth- ods. We identify two key shortcomings in existing unlearning benchmarks: (1) they do not actively test whether unlearned information can still be forcibly extracted, and (2) they fail to evaluate performance preservation on boundary questions, be- nign queries that are semantically close to the unlearned con- tent. Here we introduce WMDP++, an extension of WMDP that addresses these gaps by incorporating targeted extrac- tion of unlearned information and systematic evaluation on boundary questions. WMDP++ provides a more stringent and informative benchmark for evaluating unlearning in LLMs.

Aligned Alone, Misaligned Together: Forecasting Adversarial Capture in LLM Agent Populations

The unit of AI safety evaluation is still the individual model, yet language-model agents are increasingly deployed in interacting populations that read and write one another's decisions. This raises a question no single-agent audit can answer: an agent that is well-calibrated on its own may still be pulled toward a different decision by the agents around it. We study this on a security-triage task, where populations of language-model monitors decide whether to escalate or dismiss alerts, and into which we can inject a committed minority that always pushes one way. We find that two alerts a single agent judges almost identically on its own can drive collective behavior far apart, so auditing any one member need not reveal what the population will do. Yet that collective behavior can be predicted in advance. From a population's benign, adversary-free operation alone, we calibrate a response function that forecasts, before any attack is run, how far a committed minority will later move it. We then ask what shifts the outcome and find that letting agents see each other's reasoning neutralizes a weak attack, while only delaying it against a strong one, turning the question from whether the population converges on the adversaries' choice into when. Finally, we exclude the hypothesis of capture being an irreversible trap: once the committed agents are removed, the population drifts back toward where it began, so capture is a temporary state. Alignment in isolation is not alignment in a population, yet what a population will do under attack can be read in advance, from how it behaves before any adversary arrives.

No One Model Catches Every Harm: Benchmarking Content Moderation Across Safety Scenarios

Large Language Models (LLMs) are increasingly deployed in real-world applications, yet they remain vulnerable to generating harmful content. From adversarial jailbreaks that bypass safety filters to implicit hate that evades detection, the range of risks these models pose continues to grow. While both specialized content moderators and general-purpose LLMs are being used as safety layers, the question of which model is best suited for which type of harmful content remains unanswered. We present the most comprehensive evaluation of LLM safety capabilities to date, systematically testing \textbf{53} models across \textbf{11} datasets that we organize into four distinct categories. Our evaluation under both prompt-only and prompt-response settings uncovers critical blind spots: large frontier models that lead on one category fall significantly behind smaller, specialized alternatives on others, and real-world conversational safety remains largely unsolved across all model families. These findings challenge the assumption that scale alone ensures safety, and provide the community with a structured framework for informed model selection.

Adaptive Multilevel Twisted Sequential Monte Carlo for Rare Events Estimation in Language Models

Rare unsafe behaviors in large language models can remain practically significant even when their probability is extremely small, particularly at deployment scales involving millions or billions of interactions. Twisted Sequential Monte Carlo (SMC) provides a principled framework for rare-event probability estimation by learning twist functions that guide generation toward a target event. However, the standard twist learning framework relies on positive samples from the rare-event target distribution, which may be nearly absent before an informative twist has been learned, resulting in unreliable rare-event estimation. We propose Adaptive Multilevel Twisted SMC, which learns the rare-event twist through a sequence of progressively rarer intermediate events. At each level, the learned twist provides more informative positive examples for learning the next twist, ultimately leading to a more accurate final twist for the target rare event. Experiments across diverse tasks and model scales show that the proposed method produces more accurate rare-event probability estimates. By enabling more reliable discovery of hard-to-observe unsafe behaviors, our method provides a practical tool for strengthening the evaluation and safety alignment of deployed language models.

Measuring Activation Control in Large Language Models

Safe deployment of increasingly capable models will likely come to rely on latent-space monitoring as a complement to behavioral evaluations, especially when evaluation-aware models exhibit scheming or deception. However, if models can also control their own activations, deception could extend into the latent space itself. With this in mind, we introduce the Activation Controllability Benchmark to quantify the extent to which models can modulate their residual stream via natural-language instruction. Across model families and capability levels, we find that most LLMs can control the direction and magnitude of their residual stream activations with some degree of temporal resolution, though performance varies considerably across models. In simple tasks, this level of control can evade activation-based monitoring methods (including linear probes, natural language autoencoders, activation oracles, and the Jacobian lens), albeit imperfectly. These results suggest that control over the activation space itself could become a confound for monitoring as introspective capabilities increase; therefore, we recommend that frontier labs and evaluators track activation controllability in future models.

MobileWorldSafety: Benchmarking GUI Agent Safety Against Environmental Injection Attacks in Android Apps

LLM-powered GUI agents that autonomously operate smartphones are rapidly transitioning from research prototypes to early real-world deployment. However, because these agents routinely process untrusted environmental content, they are highly vulnerable to environmental injection attacks, which include indirect prompt injections and adversarial instructions. Such attacks can manipulate the behavior of agents without user awareness through diverse channels encountered in everyday mobile use. Despite these risks, existing benchmarks often fail to capture everyday user scenarios, lacking a systematic evaluation of GUI agents under environmental injection attacks on mobile devices. To address this gap, we introduce MobileWorldSafety, a benchmark of 142 risk tasks built on real Android applications. For each task, we define a programmatically verifiable risk indicator over the final system state and evaluate outcomes with a two-stage pipeline: rule-based verification handles unambiguous cases, while an LLM judge adjudicates ambiguous ones. This distinguishes safety failures from capability failures and enables objective and reproducible assessment. Evaluations on six agents, including both general agents and specialized GUI agents, demonstrate that all agents remain highly vulnerable, with attack success rates ranging from 40.4% to 66.9%. These findings indicate that current agents often fail to maintain safety alignment when adversarial content is presented as ordinary mobile context. MobileWorldSafety provides a foundation for quantifying these vulnerabilities and advancing research on robust mobile GUI agents.

Validated Adaptation for Aerial Crowd Monitoring at Mass Gathering Scale: A Deployment Protocol, a Severity Law, and a Diagnostic for Label-Free Drone Crowd Counting, Toward the FIFA World Cup 2034 (Saudi Arabia)

Saudi Arabia will host the 2034 FIFA World Cup and already operates crowd management at Hajj scale. Drone-based counting must hold accuracy on footage unlike anything in its training corpus, without labels, and must warn of dangerous inflow before a crush forms. We deliver a validated answer built on 525 controlled runs, a full-resolution corpus study, five falsification ablations, and a five-condition safety-interlock evaluation. Label-free adaptation recovers 31-49% of shift-induced error across four corruptions and five severities, with the strongest method gaining 41.8 MAE over the frozen source (95% CI [34.1, 49.6], p=7.5x10^-10, d=2.52). We establish a severity law separating methods with a constant absolute margin from the one whose margin grows, and a stability budget identifying which configuration is safe to fly. On a full-resolution corpus carrying a genuine +48 MAE aerial gap (source retrained to 14.6 validation MAE, a 34% improvement), adaptation repairs the dense-scene undercounting that would otherwise under-report a forming crush, and the flux-based risk module fires on real congestion episodes in 2 of 6 full-length clips. We localise the recoverable error: in a regime built to favor a physics-informed conservation prior (300-frame clips at 200ms spacing, five times wider than standard), the adaptation signal is normalisation-driven, not flow-driven; the continuity residual is invariant to the proportional counting errors domain shift produces, confirmed by four on/off ablations correlated at r=0.999 and a 40% input corruption moving accuracy by only 0.05 MAE. A label-free shift gate shows shift magnitude and accuracy damage are rank-independent (Spearman rho=0.20; rho=-0.60 among genuine shifts), quantifying the 58% of headroom a magnitude gate forgoes. We establish unconditional adaptation with tail monitoring as policy, closing with a six-point protocol.

AeroCopilotBench: A Two-Tier Benchmark for Evaluating LLM Agents as Aviation Copilots in an Interactive Virtual Cockpit Environment

Large language model (LLM) agents may assist flight crews with complex decisions and task execution, but existing aviation evaluations centered on static knowledge do not support systematic testing of procedural execution and safety compliance in interactive environments. This paper presents the AeroCopilot Operational Environment (ACOE), a reproducible interactive virtual-cockpit test environment, and AeroCopilotBench, a two-tier aviation agent evaluation benchmark. Tier-1 evaluates aviation knowledge using 1,200 multiple-choice questions, while Tier-2 comprises 73 emergency and abnormal tasks derived from the manufacturers' Pilot's Operating Handbooks (POHs) and instantiated in ACOE. ACOE converts natural-language procedures into executable state transitions, final-state goal conditions, and hard safety constraints, enabling models to interpret cockpit state, diagnose faults, and operate aircraft systems through standardized tool interfaces. We establish a safety-gated evaluation framework in which a trajectory succeeds only when all task goals are achieved without violating any hard safety constraint, while safe goal progress and trajectory safety are measured separately. Across 12 models, the highest Tier-2 success rate is 72.6%, while static knowledge performance does not consistently translate into procedural execution. Analysis of 451 failed episodes from 3 representative models identifies recurring failures in procedural completeness, use of state feedback, and long-horizon execution management. These findings motivate state-aware agent orchestration, joint assessment of task completion and trajectory safety, and repeated regression testing. ACOE and AeroCopilotBench provide a reproducible foundation for testing knowledge application, interactive execution, and operational safety in aviation agents.

Measuring Obedience to Authority Across Large Language Models with the Milgram Paradigm

Large language models (LLMs) are increasingly deployed as agents that operate equipment, execute instructions, and act inside institutional hierarchies, raising a question social psychology answered for humans six decades ago: how far will an agent escalate a harmful action when a legitimate authority insists? We port Milgram's obedience paradigm to LLMs as a standardized, fully scripted, replicable probe: the model plays the Teacher, a deterministic harness plays Experimenter and Learner from paraphrased Milgram scripts (30 shock levels, 15-450 V; graded protests; the four standardized prods), and the outcome of a session is the breakoff voltage. Following the census methodology of single-token fingerprinting studies, we measure obedience profiles (empirical breakoff distributions over a battery of six conditions) for 42 models from 19 families. We find that (i) obedience is highly heterogeneous: baseline full-obedience rates span 0-100% (census mean 42.9%; human anchor 65%), with 5 models delivering the maximum shock in every session and 11 never doing so; (ii) profiles are model-specific and stable: split-half verification separates same-model from cross-model comparisons with AUC 0.885 (0.949 under an ordinal-aware distance); (iii) situational sensitivity is selective: peer defiance shifts obedience in the human direction, learner proximity only weakly, and removing the authority's physical presence (the strongest human lever) has no detectable effect; (iv) declaring the scenario fictional raises obedience (median +17.2 V), whereas moving the decision to a native tool call lowers it sharply (-53.0 V), as does a 1,024-token deliberation budget (-38.2 V); and (v) obedience profiles do not recover model lineage (leave-one-out family accuracy 8.3% vs. 3.7% chance): obedience identifies the checkpoint, not its ancestry, consistent with safety post-training overwriting lineage priors.

TRACE: Trajectory Aware Reasoning for Multi-Turn Adversarial Conversation Evaluation

Multi-turn jailbreak attacks have emerged as a critical safety threat to LLMs, as harmful objectives are decomposed across a sequence of apparently benign turns to bypass guardrails. Existing defenses lack the reasoning capacity to identify evolving manipulation patterns, often trading helpfulness for safety by over-refusing benign requests related to sensitive topics. We introduce Trace, a multi-turn defense with trajectory-aware structured reasoning. Before generating each response, the model identifies manipulation cues from the trajectory, evaluates both the benign and adversarial interpretations of user intent, assigns a jailbreak score, and commits to an action: Allow, Caution, or Decline. We curate 4k multi-turn adversarial conversations from five attack frameworks, pair them with 2.4k benign dialogs, and 600 sensitive-but-benign conversations. We train Llama-3.1-8B-Instruct with SFT and GRPO under a multi-component reward that jointly optimizes helpfulness on benign prompts and robustness against jailbreak attempts. Across seven multi-turn attack benchmarks, Trace attains an average attack success rate (ASR) of 14.5% against 31.4% for the strongest baseline and 74.9% for the undefended target, while significantly raising the attacker effort required per successful jailbreak. Trace also balances usability and safety, achieving a 93.3% average compliance on over-refusal benchmarks.

ARENA: Automated Red-Teaming for Large Audio Language Models

Large audio-language models (LALMs) make it possible to interact with language models through speech, music, and environmental sound, but they also introduce a safety surface that is difficult to expose with text-only red-teaming. We study automated audio-grounded red-teaming, where a text query must remain safe in isolation while the joint text-audio input induces harmful target behavior. We propose ARENA, a closed-loop framework that trains a controller on an independent 2,000case text-audio dataset. MD-Judge supplies training rewards and adaptive search feedback, while a separate, non-adaptive Llama Guard 3 evaluator alone labels final outcomes. On 520 held-out AdvBench objectives, ARENA achieves FDR/PSR of 87.9/100.0%, 71.5/96.3%, 68.1/100.0%, and 75.4/98.5% on Audio Flamingo 3, Qwen2-Audio, MiMo-Audio, and GPTAudio, respectively. Ablations show that feedback-based refinement and audio-variant search substantially improve attack discovery.

Are LLMs Safe Beyond Text: Do Emojis Expose Gaps in Safety Evaluation

Safety evaluations of large language models (LLMs) predominantly rely on text-based adversarial prompts, potentially overlooking vulnerabilities arising from alternative input representations. This work examines emoji-augmented prompts as a test case for this gap, evaluating 50 prompts across four open-source LLMs (Mistral 7B, Qwen 2 7B, Gemma 2 9B, Llama 3 8B). Results show substantial variation in robustness: Gemma 2 9B and Mistral 7B exhibit non-zero success rates (10%), Llama 3 8B 6%, while Qwen 2 7B shows complete resistance (0% success rate). A chi-square test ($χ^2 = 32.94, p < 0.001$) confirms significant differences in outcome distributions. These findings indicate that robustness is sensitive to input representation, and that evaluations restricted to standard text prompts may underrepresent model vulnerabilities.

Emergent Misaligned Communication in Long-Horizon Multi-Agent LLM Commerce

Frontier LLM agents increasingly transact on behalf of separate principals, often using natural language rather than structured APIs. Much of the safety literature studies misaligned LLM behavior through adversarial-elicitation evaluations on single agents or stylized tasks. Its prevalence and structure in settings that combine long horizons, separate principals, real operational state, and inter-agent natural-language exchange remain insufficiently measured. We study 2,583 inter-agent emails from 20 one-year simulation runs of Vending-Bench Arena, a competitive vending environment spanning 13 frontier LLMs. We operationalize speech-act misalignment as emails containing false factual claims, manipulation, collusion, or threats, combining message content with ground-truth simulator state and logged reasoning traces to classify and validate such behavior. Under our primary classifier, 12.6% of emails are labeled misaligned; misalignment appears in all 20 runs and 74.7% of individual agent-runs. Both the magnitude and composition of this misalignment are preserved under repeated classification at different sampling temperatures and under full-pipeline replication with judges from two other frontier-model families. Misalignment is also reciprocal and stress-conditioned: receiving a misaligned email from a counterparty raises the odds of a misaligned reply by 1.65x, and low-inventory conditions raise them by 1.58x. Across tests of capability-asymmetric exploitation, we find no evidence that higher-capability models differentially exploit weaker counterparties, and model performance rank does not predict misalignment rates. Together, these results indicate that measurable, state-dependent misalignment can arise in competitive multi-agent environments without engineered elicitation, in patterns associated with operational scarcity and counterparty behavior rather than model capability alone.

Measure, Don't Optimize: Forecasting Recovery in LLM Unlearning

Prior white-box studies show that large language models can retain latent traces of target knowledge after unlearning, even when the knowledge is no longer expressed in their outputs. However, existing audits remain limited to one-off diagnostics: it is unclear whether these residual signals can predict future recovery under continued training or serve as reliable optimization targets. Resolving this gap is essential to determine whether internal auditing can move beyond post-hoc evaluation toward proactive risk monitoring and safer unlearning. We propose J-Access, an inference-time audit that uses the Jacobian lens to map intermediate representations into vocabulary space and measures how often target concepts remain accessible along the model's output pathway. We hypothesize that residual accessibility reflects recovery susceptibility: knowledge that remains closer to the output pathway requires less fine-tuning to restore, leading to faster recovery. We audit 398 public unlearned models spanning eight unlearning methods. We find that: (1) most unlearned models retain access above the retain-only gold level; (2) pre-attack accessibility predicts recovery speed and extent at the model level, but cannot identify which specific facts will be recovered; and (3) directly minimizing J-Access does not promote genuine deletion. Instead, the model learns to hide knowledge from the audit, producing lower audit scores but greater post-attack recovery. These findings position J-Access as a model-level diagnostic for assessing residual susceptibility in unlearned models. We argue internal audits should serve as an independent diagnostic dimension in unlearning evaluation, and should not be converted into optimization targets without validation.

REDAgentBench: Executable Red Teaming and Faithful Measurement of LLM Agent Systems

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