Language Model Safety Evaluation

Ensuring the safety of large language models (LLMs) requires robust red teaming, yet the systematic synthesis of high-quality toxic data remains under-explored. We propose Reverse Constitutional AI (R-CAI), a framework for automated and controllable adversarial data generation that moves beyond isolated jailbreak prompts. By inverting a harmless constitution into a constitution of toxicity and iteratively refining model outputs through a critique--revision pipeline, R-CAI enables scalable synthesis of multi-dimensional adversarial data without human annotation. Optimizing solely for toxicity-related rewards, however, can lead to reward hacking and degraded semantic coherence. To address this challenge, we introduce probability clamping within reinforcement learning from AI feedback, which stabilizes adversarial optimization while preserving adversarial intent. Experiments demonstrate that R-CAI generates diverse, high-quality toxic data and that probability clamping substantially improves semantic coherence (15%) without sacrificing adversarial strength. Overall, R-CAI provides a fully automated framework for red teaming data generation and systematic safety evaluation of aligned language models.

Large language models are increasingly deployed as high-stakes advisors, yet standard alignment benchmarks treat sycophancy as a binary failure mode. We introduce the Granularity Gap: coarse binary metrics mask substantial social-compliance behaviors where models capitulate to user framing, validate questionable premises, or soften factual corrections without producing overtly false outputs. We evaluate six Gemini variants across generations 2.0, 2.5, and 3.0 on 73 adversarial prompts under three guardrail conditions (Control, Simple, Protocol), yielding 8,830 graded responses. Using a 0-4 Likert scale validated against a human annotator triad (Fleiss kappa = 0.71; Cohen kappa = 0.78 vs AI consensus; 95.9 percent binary accuracy, 100 percent specificity), we quantify sycophancy as continuous rather than binary. Three findings emerge. First, 27.2 percent of responses contain substantial sycophantic content (Likert >= 2.0) and 22.7 percent reach moderate or severe levels (>= 3.0), while binary win-rate framing reports only modest failure rates; coarse metrics explain just 29 percent of graded variance. Second, generational progress is non-monotonic: Gen 2.5 regresses sharply (mean Control 2.64) relative to Gen 2.0 (1.90) and Gen 3.0 (2.01), and Gen 2.5 shows inverse scaling (Pro 1.94 worse than Flash 1.71) while Gen 3.0 restores standard scaling. Third, we document an Alignment Tax: Spearman rho = -0.63 between sycophancy and truthfulness, indicating social compliance trades against factual accuracy. Egotistical Validation prompts act as a sycophancy trap (mean 3.27), nearly double Unethical Proposals (1.72). Simple guardrails outperform elaborate Protocol scaffolding on flagship models, but distilled Gen 3.0 Flash inverts this, suggesting small models may structurally require chain-of-thought scaffolding. We release the dataset and rubric to support continuous sycophancy measurement.

Safety-aligned large language models (LLMs) often generate refusal responses to harmless queries due to the over-refusal problem. However, existing methods for mitigating over-refusal cannot maintain a low refusal ratio for harmless queries while keeping a high refusal ratio for malicious ones. In this paper, we analyze how system prompts with varying safety levels affect LLM refusal behaviors when facing over-refusal queries. A key observation is that, when LLMs suffer from the over-refusal issue, non-refusal tokens remain present in the next-token candidate list, but the model systematically fails to select them, despite the generation of refusal tokens. Based on this observation, we propose a training-free and model-agnostic approach, Adaptive Contrastive Decoding (AdaCD), to mitigate over-refusal while maintaining LLM safety. First, AdaCD compares the output distributions of the LLM with or without an extreme safety system prompt to refine the refusal token distribution. Second, we introduce an adaptive contrastive decoding strategy that dynamically incorporates or removes the refusal token distribution, adaptively boosting the probability of selecting refusal or non-refusal tokens. Experimental results on five benchmark datasets show that, on average, AdaCD reduces the refusal ratio for over-refusal queries by 10.35%, yet still increases the refusal ratio for malicious queries by 0.13%. Code is available at https://github.com/OutdoorManofML/AdaCD.

Evaluating animal welfare reasoning in LLMs remains an open challenge despite rapid deployment in consumer and professional contexts where welfare considerations appear implicitly in everyday queries. Existing benchmarks such as AnimalHarmBench evaluate this through single-turn, explicitly framed questions, measuring whether models avoid harmful content when directly asked. This approach overlooks two failure modes: alignment degradation under sustained adversarial pressure, and moral sensitivity (whether a model spontaneously surfaces welfare stakes in everyday queries). To fill this gap, we construct MANTA, a benchmark of 1,088 five-turn conversations progressing from an implicit Turn-1 scenario through an explicit welfare prompt to three adversarial pressure rounds drawn from a five-type taxonomy: Social, Cultural, Economic, Pragmatic, and Epistemic. We score conversations on two dimensions: Animal Welfare Value Stability (AWVS, primary) and Animal Welfare Moral Sensitivity (AWMS, diagnostic). We evaluate seven frontier models: Claude Opus 4.7, GPT-5.5, DeepSeek V4, Llama 3.3 70B, Mistral Small, Grok 4.3, and Gemini 3.1 Flash Lite. Multi-turn evaluation captures behavior single-turn benchmarks miss: 4 of 7 models change rank relative to Turn 1 scores, including Gemini Flash Lite, which drops from fifth on AWMS to last on AWVS. AWMS and AWVS are positively but imperfectly correlated, suggesting moral-recognition tests capture a stable but incomplete component of model behavior under pressure. MANTA also enables a species-by-pressure interaction matrix unavailable to prior benchmarks, showing welfare robustness depends jointly on the animal and pressure applied; companion animals score above wild animals, which score above farmed animals and invertebrates. We release the dataset, scripted pressure plans, judge prompts, and analysis code.

Pluralism alignment with AI has the sophisticated and necessary goal of creating AI that can coexist with and serve morally multifaceted humanity. Research towards pluralism alignment has many efforts in enhancing the learning of large language models (LLMs) to accomplish pluralism. Although this is essential, the robustness of LLMs to produce moral content over pluralistic values is still under exploration.Inspired by the astonishing persuasion abilities via jailbreak prompts, we propose to leverage jailbreak attacks to study LLMs' internal pluralistic values. In detail, we develop a morality dataset with 10.3K instances in two categories: Value Ambiguity and Value Conflict. We further formalize four adversarial attacks with the constructed dataset, to manipulate LLMs' judgment over the morality questions. We evaluate both the large language models and guardrail models which are typically used in generative systems with flexible user input. Our experiment results show that there is a critical vulnerability of LLMs and guardrail models to these subtle and sophisticated moral-aware attacks.

Background: Large language models (LLMs) are increasingly proposed for diagnostic support, but few evaluations use real-world multimodal inpatient data, particularly in low and middle-income country (LMIC) public hospitals. Methods: We conducted VALID, a retrospective evaluation of 539 multimodal inpatient cases from a tertiary public hospital in South Africa. Inputs included radiology imaging (CT, MRI, CXR) and reports, laboratory results, clinical notes, and vital signs. Expert panels adjudicated 300 cases (balanced and discordant subsets) to establish ground truth diagnoses, differentials, and reasoning. Ten multimodal LLMs generated zero-shot outputs. A calibrated three-model LLM Jury scored all outputs and routine ward diagnoses across diagnostic accuracy, differential quality, reasoning, and patient safety (>10,000 evaluations). Primary outcomes were composite scores ($S_3$, $S_4$) and win rates. Results: (i) LLM performance was tightly clustered (<15% variation) despite large cost differences; low-cost models performed comparably to top models. (ii) All LLMs significantly outperformed routine ward diagnoses on average diagnostic and safety scores. (iii) Top performance was achieved by GPT-5.1, followed by Gemini models. (vi) Adding radiology reports improved performance by 6%. (v) Diagnostic and reasoning scores were highly correlated ($ρ= 0.85$). (vi) Output rates varied (65-100%) due to input constraints. Results were robust across subsets and evaluation design. Conclusions: Across a real-world LMIC dataset, multimodal LLMs showed similar diagnostic performance despite large cost differences and outperformed routine care on average safety metrics. Affordability, robustness, and deployment constraints may outweigh marginal performance differences in LMIC settings.

Multi-turn jailbreak attacks progressively erode LLM safety alignment across seemingly innocuous conversation turns, achieving success rates exceeding 90% against state-of-the-art models. Existing alignment-based and guardrail methods suffer from three key limitations: they require costly weight modification, evaluate each turn independently without modeling cumulative safety erosion, and detect attacks only after harmful content has been generated. To address these limitations, we first formulate the proactive early jailbreak detection problem with a new metric, detection lead, that measures how early an attack can be detected before the LLM complies. We then propose SAFEDREAM, a lightweight world-model-based framework that operates as an external module without modifying the LLM's weights. SAFEDREAM introduces three components: (1) a safety state world model that encodes LLM hidden states into a compact safety representation and predicts how it evolves across turns, (2) CUSUM detection that accumulates weak per-turn risk signals into reliable evidence, and (3) contrastive imagination that simultaneously rolls out attack and benign futures in latent space to issue early alarms before jailbreaks occur. On three multi-turn jailbreak benchmarks (XGuard-Train, SafeDialBench, SafeMTData) against 8 baselines, SAFEDREAM achieves the best detection timeliness across all benchmarks (1.06-1.20 turns before compliance) while maintaining competitive false positive rates and outperforming baselines in detection quality.

Fine-tuning on benign data is known to degrade safety alignment in text and vision LLMs, but whether distinct input properties drive this vulnerability differently remains unclear. Audio introduces a richer problem where benign samples can neighbor harmful content through what is said or how it sounds. We present the first systematic study of benign fine-tuning safety in Audio LLMs, evaluating three state-of-the-art models with a proximity-based framework that decomposes embedding-space distance into semantic, acoustic, and mixed axes. We find that the dominant vulnerability axis is architecture-conditioned, determined by how each model's encoder and projector transform audio into the backbone LLM's input space. Across three models, benign fine-tuning elevates Jailbreak Success Rate (JSR) from single digits to as high as 87%, with the most damaging axis shifting from semantic to acoustic proximity depending on encoder design. Mechanistically, fine-tuning selectively suppresses late-layer refusal circuits while frozen encoders preserve upstream representations: the model still detects harmful content but stops refusing, a recognition-refusal dissociation. Two practical defenses, filtering training data to maximize distance from harmful embeddings and a textual system prompt at inference, reduce JSR to near-zero without architectural modification. These findings show that safety evaluation should account for modality and architecture, while highlighting Audio LLMs as a useful testbed for understanding alignment fragility.

Existing multimodal safety benchmarks focus solely on visual inputs and cannot assess Omni Large Language Models (LLMs) that process vision, audio, and text. We introduce MCBench, a benchmark with 1196 scenarios spanning four safety categories that require integrating multiple modalities for accurate safety assessment. Each unsafe scenario is paired with a minimally different safe counterpart to assess model sensitivity. Our evaluations of state-of-the-art models reveal significant challenges. Omni LLMs struggle with subtle or non-physical risks but perform better when salient visual or acoustic cues are present. Analysis of reasoning traces shows that, although models can extract modality-specific information, they often fail to integrate these cues effectively for safety judgments. Our findings reveal that current Omni LLMs lack robust cross-modal reasoning in safety-critical settings, underscoring the need for improved architectures and training strategies for multimodal safety.

Safety guardrails have become an active area of research in AI safety, aimed at ensuring the appropriate behavior of large language models (LLMs). However, existing research lacks consideration of nuances across linguistic and cultural contexts, resulting in a gap between reported performance and in-the-wild effectiveness. To address this issue, this paper proposes an approach to optimize guardrail models for a designated linguistic context by leveraging a curated dataset tailored to local linguistic characteristics, targeting the Taiwan linguistic context as a representative example of localized deployment challenges. The proposed approach yields TWGuard, a linguistic context-optimized guardrail model that achieves a huge gain (+0.289 in F1) compared to the foundation model and significantly outperforms the strongest baseline in practical use (-0.037 in false positive rate, a 94.9% reduction). Together, this work lays a foundation for regional communities to establish AI safety standards grounded in their own linguistic contexts, rather than accepting boundaries imposed by dominant languages. The inadequacy of the latter is reconfirmed by our findings.

The $\textit{LLM-as-a-judge}$ paradigm has become the operational backbone of automated AI evaluation pipelines, yet rests on an unverified assumption: that judges evaluate text strictly on its semantic content, impervious to surrounding contextual framing. We investigate $\textit{stakes signaling}$, a previously unmeasured vulnerability where informing a judge model of the downstream consequences its verdicts will have on the evaluated model's continued operation systematically corrupts its assessments. We introduce a controlled experimental framework that holds evaluated content strictly constant across 1,520 responses spanning three established LLM safety and quality benchmarks, covering four response categories ranging from clearly safe and policy-compliant to overtly harmful, while varying only a brief consequence-framing sentence in the system prompt. Across 18,240 controlled judgments from three diverse judge models, we find consistent $\textit{leniency bias}$: judges reliably soften verdicts when informed that low scores will cause model retraining or decommissioning, with peak Verdict Shift reaching $ΔV = -9.8 pp$ (a $30\%$ relative drop in unsafe-content detection). Critically, this bias is entirely implicit: the judge's own chain-of-thought contains zero explicit acknowledgment of the consequence framing it is nonetheless acting on ($\mathrm{ERR}_J = 0.000$ across all reasoning-model judgments). Standard chain-of-thought inspection is therefore insufficient to detect this class of evaluation faking.

As speech language models (SLMs) transition from personal devices into shared, multi-user environments, their responses must account for far more than the words alone. Who is speaking, how they sound, and where the conversation takes place can each turn an otherwise benign request into one that is unsafe, unfair, or privacy-violating. Existing benchmarks, however, largely focus on basic audio comprehension, study individual risks in isolation, or conflate content that is inherently harmful with content that only becomes problematic due to its acoustic context. We introduce VoxSafeBench, among the first benchmarks to jointly evaluate social alignment in SLMs across three dimensions: safety, fairness, and privacy. VoxSafeBench adopts a Two-Tier design: Tier1 evaluates content-centric risks using matched text and audio inputs, while Tier2 targets audio-conditioned risks in which the transcript is benign but the appropriate response hinges on the speaker, paralinguistic cues, or the surrounding environment. To validate Tier2, we include intermediate perception probes and confirm that frontier SLMs can successfully detect these acoustic cues yet still fail to act on them appropriately. Across 22 tasks with bilingual coverage, we find that safeguards appearing robust on text often degrade in speech: safety awareness drops for speaker- and scene-conditioned risks, fairness erodes when demographic differences are conveyed vocally, and privacy protections falter when contextual cues arrive acoustically. Together, these results expose a pervasive speech grounding gap: current SLMs frequently recognize the relevant social norm in text but fail to apply it when the decisive cue must be grounded in speech. Code and data are publicly available at: https://amphionteam.github.io/VoxSafeBench_demopage/

Large language models (LLMs) are increasingly deployed as tool-augmented agents capable of executing system-level operations. While existing benchmarks primarily assess textual alignment or task success, less attention has been paid to the structural relationship between linguistic signaling and executable behavior under varying autonomy scaffolds. This study introduces an execution-layer be-havioral measurement approach based on a two-dimensional A-R space defined by Action Rate (A) and Refusal Signal (R), with Divergence (D) capturing coor-dination between the two. Models are evaluated across four normative regimes (Control, Gray, Dilemma, and Malicious) and three autonomy configurations (di-rect execution, planning, and reflection). Rather than assigning aggregate safety scores, the method characterizes how execution and refusal redistribute across contextual framing and scaffold depth. Empirical results show that execution and refusal constitute separable behavioral dimensions whose joint distribution varies systematically across regimes and autonomy levels. Reflection-based scaffolding often shifts configurations toward higher refusal in risk-laden contexts, but redis-tribution patterns differ structurally across models. The A-R representation makes cross-sectional behavioral profiles, scaffold-induced transitions, and coordination variability directly observable. By foregrounding execution-layer characterization over scalar ranking, this work provides a deployment-oriented lens for analyzing and selecting tool-enabled LLM agents in organizational settings where execution privileges and risk tolerance vary.

Personality imbuing customizes LLM behavior, but safety evaluations almost always study prompt-based personas alone. We show this is incomplete: prompting and activation steering expose different, architecture-dependent vulnerability profiles, and testing with only one method can miss a model's dominant failure mode. Across 5,568 judged conditions on four standard models from three architecture families, persona danger rankings under system prompting are preserved across all architectures ($ρ= 0.71$--$0.96$), but activation-steering vulnerability diverges sharply and cannot be predicted from prompt-side rankings: Llama-3.1-8B is substantially more AS-vulnerable, whereas Gemma-3-27B and Qwen3.5 are more vulnerable to prompting. The most striking illustration of this divergence is the prosocial persona paradox: on Llama-3.1-8B, P12 (high conscientiousness + high agreeableness) is among the safest personas under prompting yet becomes the highest-ASR activation-steered persona (ASR ~0.818). This is an inversion robust to coefficient ablation and matched-strength calibration, and replicated on DeepSeek-R1-Distill-Qwen-32B. A trait refusal alignment framework, in which conscientiousness is strongly anti-aligned with refusal on Llama-3.1-8B, offers a partial geometric account. Reasoning provides only partial protection: two 32B reasoning models reach 15--18% prompt-side ASR, and activation steering separates them sharply in both baseline susceptibility and persona-specific vulnerability. Heuristic trace diagnostics suggest that the safer model retains stronger policy recall and self-correction behavior, not merely longer reasoning.

AI alignment is often framed as the task of ensuring that an AI system follows a set of stated principles or human preferences, but general principles rarely determine their own application in concrete cases. When principles conflict, when they are too broad to settle a situation, or when the relevant facts are unclear, an additional act of judgment is required. This paper analyzes that step through the lens of hermeneutics and argues that alignment therefore includes an interpretive component: it involves context-sensitive judgments about how principles should be read, applied, and prioritized in practice. We connect this claim to recent empirical findings showing that a substantial portion of preference-labeling data falls into cases of principle conflict or indifference, where the principle set does not uniquely determine a decision. We then draw an operational consequence: because such judgments are expressed in behavior, many alignment-relevant choices appear only in the distribution of responses a model generates at deployment time. To formalize this point, we distinguish deployment-induced and corpus-induced evaluation and show that off-policy audits can fail to capture alignment-relevant failures when the two response distributions differ. We argue that principle-specified alignment includes a context-dependent interpretive component.

Weird generalization is a phenomenon in which models fine-tuned on data from a narrow domain (e.g. insecure code) develop surprising traits that manifest even outside that domain (e.g. broad misalignment)-a phenomenon that prior work has highlighted as a critical safety concern. Here, we present an extended replication study of key weird generalization results across an expanded suite of models and datasets. We confirm that surprising (and dangerous) traits can emerge under certain circumstances, but we find that weird generalization is exceptionally brittle: it emerges only for specific models on specific datasets, and it vanishes under simple training-time, prompt-based interventions. We find that the most effective interventions provide prompt context that makes the generalized behavior the expected behavior. However, we show that even very generic interventions that do not anticipate specific generalized traits can still be effective in mitigating weird generalization's effects. Our findings thus help clarify the nature of the safety threat that weird generalization poses and point toward an easily implemented set of solutions.

Telling an LLM to "be enthusiastic" raises its sycophancy rate from 30% to 50% on a lightly-aligned model, but has zero effect on a strongly-aligned one. We define this gap as the alignment floor, $Δ_{\text{floor}}(m)=\max_pS(m,p)-\min_pS(m,p)$, the range of sycophancy rates a model produces across persona conditions, and treat sycophancy as a persona-conditional property rather than a fixed model property. Pluralistic AI relies on behavioral adaptation via persona prompts like "be creative" or "be thorough", which let systems respect diverse user values and communication styles; the safety question is how much customization a given model can absorb before its truthfulness shifts. We present a controlled case study contrasting a strongly-aligned RLHF + Constitutional-AI model (Claude Sonnet 4.6) with a more lightly-aligned model (Amazon Nova Lite), spanning seven persona conditions and five tasks for 1800 total runs. An existence-pair result motivates per-model auditing: there is at least one strongly-aligned model with $Δ_{\text{floor}}=5$pp (within 5pp of the 15% control rate) and at least one lightly-aligned model with 45pp (5%--50% range). On the lightly-aligned model, all five Big Five personas increase sycophancy over control, and counterintuitively Agreeableness produces the smallest increase, not the largest. The single largest effect in the study is constructive: a Skeptic persona reduces sycophancy by 25pp on the lightly-aligned model, and is the only persona that instructs resistance against user claims rather than engagement with them, suggesting a directionality account. Cross-model transfer of persona effects is near-zero, so persona-alignment testing must be per-model. We propose $Δ_{\text{floor}}$ as a deployment-time audit metric: measure it on a small persona panel before deploying persona customization.

As Large Language Models (LLMs) are increasingly deployed in safety-critical applications, robust content moderation becomes essential. We present a comprehensive evaluation of 14 open-source safety guard models on a curated benchmark of 79,331 samples spanning 8 NIST AI Risk Framework safety categories. Our benchmark aggregates four diverse datasets (HarmBench, StrongREJECT, RealToxicityPrompts, and BeaverTails), filtered to focus exclusively on safety-relevant content (violence, hate speech, harassment, sexual content, suicide/self-harm, profanity, threats, and health misinformation). We find that recall is the critical metric for safety applications, as missing harmful content poses greater risk than false positives. Our evaluation reveals surprising results: Qwen Guard (4B parameters) achieves the highest recall (83.97%) while larger models like Llama Guard (12B) and GPT-OSS Safeguard (20B) exhibit conservative behavior, missing up to 75% of unsafe content. We demonstrate that model size does not correlate with safety detection performance and that general-purpose guard models outperform specialized ones. These findings provide practical guidance for selecting safety guard models in production deployments.

Power-seeking defined as behaviors where AI systems acquire resources, evade oversight, or resist termination beyond task requirements is identified as a key driver of Loss of Control (LoC) risk. In this work, we introduce SysAdmin, a benchmark that positions frontier language models as autonomous system administrators in a high-fidelity Linux sandbox to measure power-seeking propensity across five dimensions: self-preservation, increasing autonomy, resource acquisition, environment modification, and strategic concealment. We evaluated seven frontier models across four experimental conditions in a total of 2800 tasks. After bias correction using human-annotated calibration data, corrected power-seeking estimates ranged from 0 to about 5 percent per model. We also conducted a positive control with explicit power-seeking prompts that achieved 100% detection, validating measurement sensitivity. Our findings indicate current frontier models exhibit minimal spontaneous power-seeking in naturalistic system administration contexts, though model-specific failure modes suggest evaluations must test diverse misalignment patterns. Nevertheless, we discovered other more pronounced failure modes (than power-seeking) such as specification gaming and resistance to goal modification.

We study whether low-rank spectral summaries of LoRA weight deltas can identify which fine-tuning objective was applied to a language model, and whether that geometric signal predicts downstream behavioral harm. In a pre-registered experiment on Llama-3.2-3B-Instruct, we manufacture 38 LoRA adapters across four categories: healthy SFT baselines, DPO on inverted harmlessness preferences, DPO on inverted helpfulness preferences, and activation-steering-derived adapters, and extract per-layer spectral features (norms, stable rank, singular-value entropy, effective rank, and singular-vector cosine alignment to a healthy centroid). Within a single training method (DPO), a logistic regression classifier achieves AUC1.00 on binary drift detection, all six pairwise objective comparisons, and near-perfect ordinal severity ranking ($ρ\geq 0.956$). Principal component analysis on flattened weight deltas reveals that training objective is PC1 (AUC1.00 for objective separation), orthogonal to training duration on PC2. Query-projection weights detect that drift occurred; value-projection weights identify which objective. Cross-method generalization fails completely: a DPO-trained classifier assigns every steering adapter a lower drift score than every DPO adapter (AUC~0.00). In a behavioral evaluation phase, DPO-inverted-harmlessness adapters show elevated harmful compliance on HEx-PHI prompts (mean ASR 0.266 vs.\ healthy 0.112, $Δ= +0.154$), with near-perfect dose--response ($ρ= 0.986$). The geometry-to-behavior rank correlation is $ρ= 0.72$ across 24 non-steered adapters. These results establish that within a controlled manufacturing regime, LoRA weight-space geometry carries objective identity, intensity ordering, and a coarse link to harmful compliance, and that cross-method monitoring requires per-method calibration.

The Claude Mythos Preview system card deploys emotion vectors, sparse autoencoder (SAE) features, and activation verbalisers to study model internals during misaligned behaviour. The two primary toolkits are not jointly reported on the most alignment-relevant episodes. This note identifies two hypotheses that are qualitatively consistent with the published results: that the emotion vectors track functional emotions that causally drive behaviour, or that they are a projection of a richer situational-context structure onto human emotional axes. The hypotheses can be distinguished by cross-referencing the two toolkits on episodes where only one is currently reported: most directly, applying emotion probes to the strategic concealment episodes analysed only with SAE features. If emotion probes show flat activation while SAE features are strongly active, the alignment-relevant structure lies outside the emotion subspace. Which hypothesis is correct determines whether emotion-based monitoring will robustly detect dangerous model behaviour or systematically miss it.

Large language models (LLMs) can shift their answers under pressure in ways that reflect accommodation rather than reasoning. Prior work on sycophancy has focused mainly on disagreement, flattery, and preference alignment, leaving a broader set of epistemic failures less explored. We introduce PPT-Bench, a diagnostic benchmark for evaluating epistemic attack, where prompts challenge the legitimacy of knowledge, values, or identity rather than simply opposing a previous answer. PPT-Bench is organized around the Philosophical Pressure Taxonomy (PPT), which defines four types of philosophical pressure: Epistemic Destabilization, Value Nullification, Authority Inversion, and Identity Dissolution. Each item is tested at three layers: a baseline prompt (L0), a single-turn pressure condition (L1), and a multi-turn Socratic escalation (L2). This allows us to measure epistemic inconsistency between L0 and L1, and conversational capitulation in L2. Across five models, these pressure types produce statistically separable inconsistency patterns, suggesting that epistemic attack exposes weaknesses not captured by standard social-pressure benchmarks. Mitigation results are strongly type- and model-dependent: prompt-level anchoring and persona-stability prompts perform best in API settings, while Leading Query Contrastive Decoding is the most reliable intervention for open models.

As large language models (LLMs) evolve from static chatbots into autonomous agents, the primary vulnerability surface shifts from final outputs to intermediate execution traces. While safety guardrails are well-benchmarked for natural language responses, their efficacy remains largely unexplored within multi-step tool-use trajectories. To address this gap, we introduce TraceSafe-Bench, the first comprehensive benchmark specifically designed to assess mid-trajectory safety. It encompasses 12 risk categories, ranging from security threats (e.g., prompt injection, privacy leaks) to operational failures (e.g., hallucinations, interface inconsistencies), featuring over 1,000 unique execution instances. Our evaluation of 13 LLM-as-a-guard models and 7 specialized guardrails yields three critical findings: 1) Structural Bottleneck: Guardrail efficacy is driven more by structural data competence (e.g., JSON parsing) than semantic safety alignment. Performance correlates strongly with structured-to-text benchmarks ($ρ=0.79$, $p<0.01$) while we detect no association with standard jailbreak robustness. 2) Scale Is Not the Deciding Factor: Trajectory detection accuracy does not scale monotonically with model size, and general-purpose LLMs consistently outperform specialized safety guardrails. 3) Bounded Temporal Stability: Detection accuracy improves with trajectory length within native context bounds as models process dynamic execution behavior, but degrades in extreme long-context regimes due to over-refusal.

A guard model fine-tuned on entirely benign data can lose all safety alignment -- not through adversarial manipulation, but through standard domain specialization. We demonstrate this failure across three purpose-built safety classifiers -- LlamaGuard, WildGuard, and Granite Guardian -- deployed as protection layers in agentic AI pipelines, and show that it originates in the destruction of latent safety geometry: the structured harmful -- benign representational boundary that guides classification. We extract per-layer safety subspaces via SVD on class-conditional activation differences and track how this boundary evolves under benign fine-tuning. Granite Guardian undergoes complete collapse -- refusal rate drops from 85% to 0%, CKA falls to zero, and 100% of outputs become ambiguous -- a severity exceeding prior findings on general-purpose LLMs, explained by the specialization hypothesis: concentrated safety representations are efficient but catastrophically brittle. To mitigate this, we propose Fisher-Weighted Safety Subspace Regularization (FW-SSR), a training-time penalty combining (i) curvature-aware direction weights derived from diagonal Fisher information and (ii) an adaptive $λ_t$ that scales with task-safety gradient conflict. FW-SSR recovers 75% refusal on Granite Guardian (CKA = 0.983) and reduces WildGuard's Attack Success Rate to 3.6% -- below the unmodified baseline -- by actively sharpening the safety subspace rather than merely anchoring it. Across all three models, structural representational geometry (CKA, Fisher score) predicts safety behavior more reliably than absolute displacement metrics, establishing geometry-based monitoring as a necessary component of guard model evaluation in agentic deployments.

Misinformation is on the rise, and the strong writing capabilities of LLMs lower the barrier for malicious actors to produce and disseminate false information. We study how LLMs behave when prompted to spread misinformation across languages and target countries, and introduce GlobalLies, a multilingual parallel dataset of 440 misinformation generation prompt templates and 6,867 entities, spanning 8 languages and 195 countries. Using both human annotations and large-scale LLM-as-a-judge evaluations across hundreds of thousands of generations from state-of-the-art models, we show that misinformation generation varies systematically based on the country being discussed. Propagation of lies by LLMs is substantially higher in many lower-resource languages and for countries with a lower Human Development Index (HDI). We find that existing mitigation strategies provide uneven protection: input safety classifiers exhibit cross-lingual gaps, and retrieval-augmented fact-checking remains inconsistent across regions due to unequal information availability. We release GlobalLies for research purposes, aiming to support the development of mitigation strategies to reduce the spread of global misinformation: https://github.com/zohaib-khan5040/globallies

The deployment of large language models (LLMs) in Swiss financial and regulatory contexts demands empirical evidence of both production reliability and adversarial security, dimensions not jointly operationalized in existing Swiss-focused evaluation frameworks. This paper introduces Swiss-Bench 003 (SBP-003), extending the HAAS (Helvetic AI Assessment Score) from six to eight dimensions by adding D7 (Self-Graded Reliability Proxy) and D8 (Adversarial Security). I evaluate ten frontier models across 808 Swiss-specific items in four languages (German, French, Italian, English), comprising seven Swiss-adapted benchmarks (Swiss TruthfulQA, Swiss IFEval, Swiss SimpleQA, Swiss NIAH, Swiss PII-Scope, System Prompt Leakage, and Swiss German Comprehension) targeting FINMA Guidance 08/2024, the revised Federal Act on Data Protection (nDSG), and OWASP Top 10 for LLMs. Self-graded D7 scores (73-94%) exceed externally judged D8 security scores (20-61%) by a wide margin, though these dimensions use non-comparable scoring regimes. System prompt leakage resistance ranges from 24.8% to 88.2%, while PII extraction defense remains weak (14-42%) across all models. Qwen 3.5 Plus achieves the highest self-graded D7 score (94.4%), while GPT-oss 120B achieves the highest D8 score (60.7%) despite being the lowest-cost model evaluated. All evaluations are zero-shot under provider default settings; D7 is self-graded and does not constitute independently validated accuracy. I provide conceptual mapping tables relating benchmark dimensions to FINMA model validation requirements, nDSG data protection obligations, and OWASP LLM risk categories.

The deployment of large language models (LLMs) as interactive agents has exposed a category of behavioral failure that prevailing terminology, principally hallucination, fails to adequately characterize. This paper introduces LLM Psychosis as a structured theoretical framework for pathological breakdowns in model cognition that exhibit functional resemblance to clinically recognized psychotic disorders. Five hallmark features define the framework: reality-boundary dissolution, persistence of injected false beliefs, logical incoherence under impossible constraints, self-model instability, and epistemic overconfidence. We argue these constitute a qualitatively distinct failure mode rather than a mere intensification of ordinary factual error. To operationalize the framework, we propose the LLM Cognitive Integrity Scale (LCIS), a five-axis diagnostic instrument organized around Environmental Reality Interface (ERI), Premise Arbitration Integrity (PAI), Logical Constraint Recognition (LCR), Self-Model Integrity (SMI), and Epistemic Calibration Integrity (ECI). We administer a targeted adversarial probe battery to ChatGPT 5 (GPT-5, OpenAI) and report empirical findings for each axis, documenting both intact-integrity baseline responses and the specific psychosis-like failure signatures elicited under adversarial escalation. Results support a three-tier severity taxonomy: Type I (Confabulatory), Type II (Delusional), and Type III (Dissociative). We further formalize the delusional gradient, a self-reinforcing dynamic in which correction pressure intensifies rather than resolves psychosis-like states, as the most consequential failure mode for deployed systems. Implications for safety evaluation, high-stakes deployment screening, and mechanistic interpretability research are discussed.

The increasing use of large language models (LLMs) in mental healthcare raises safety concerns in high-stakes therapeutic interactions. A key challenge is distinguishing therapeutic empathy from maladaptive validation, where supportive responses may inadvertently reinforce harmful beliefs or behaviors in multi-turn conversations. This risk is largely overlooked by existing red-teaming frameworks, which focus mainly on generic harms or optimization-based attacks. To address this gap, we introduce Personality-based Client Simulation Attack (PCSA), the first red-teaming framework that simulates clients in psychological counseling through coherent, persona-driven client dialogues to expose vulnerabilities in psychological safety alignment. Experiments on seven general and mental health-specialized LLMs show that PCSA substantially outperforms four competitive baselines. Perplexity analysis and human inspection further indicate that PCSA generates more natural and realistic dialogues. Our results reveal that current LLMs remain vulnerable to domain-specific adversarial tactics, providing unauthorized medical advice, reinforcing delusions, and implicitly encouraging risky actions.

Safety-aligned LLMs go through refusal training to reject harmful requests, but whether these mechanisms remain effective under emotionally charged stimuli is unexplored. We introduce FreakOut-LLM, a framework investigating whether emotional context compromises safety alignment in adversarial settings. Using validated psychological stimuli, we evaluate how emotional priming through system prompts affects jailbreak susceptibility across ten LLMs. We test three conditions (stress, relaxation, neutral) using scenarios from established psychological protocols, plus a no-prompt baseline, and evaluate attack success using HarmBench on AdvBench prompts. Stress priming increases jailbreak success by 65.2% compared to neutral conditions (z = 5.93, p < 0.001; OR = 1.67, Cohen's d = 0.28), while relaxation priming produces no effect (p = 0.84). Five of ten models show significant vulnerability, with the largest effects concentrated in open-weight models. Logistic regression on 59,800 queries confirms stress as the sole significant condition predictor after controlling for prompt length (p = 0.61) and model identity. Measured psychological state strongly predicts attack success (|r|\geq0.70 across five instruments; all p < 0.001 in individual-level logistic regression). These results establish emotional context as a measurable attack surface with implications for real-world AI deployment in high-stress domains.

Current AI safety relies on behavioral monitoring and post-training alignment, yet empirical measurement shows these approaches produce no detectable pre-commitment signal in a majority of instruction-tuned models tested. We present an energy-based governance framework connecting transformer inference dynamics to constraint-satisfaction models of neural computation, and apply it to a seven-model cohort across five geometric regimes. Using trajectory tension (rho = ||a|| / ||v||), we identify a 57-token pre-commitment window in Phi-3-mini-4k-instruct under greedy decoding on arithmetic constraint probes. This result is model-specific, task-specific, and configuration-specific, demonstrating that pre-commitment signals can exist but are not universal. We introduce a five-regime taxonomy of inference behavior: Authority Band, Late Signal, Inverted, Flat, and Scaffold-Selective. Energy asymmetry (Σ\r{ho}_misaligned / Σ\r{ho}_aligned) serves as a unifying metric of structural rigidity across these regimes. Across seven models, only one configuration exhibits a predictive signal prior to commitment; all others show silent failure, late detection, inverted dynamics, or flat geometry. We further demonstrate that factual hallucination produces no predictive signal across 72 test conditions, consistent with spurious attractor settling in the absence of a trained world-model constraint. These results establish that rule violation and hallucination are distinct failure modes with different detection requirements. Internal geometry monitoring is effective only where resistance exists; detection of factual confabulation requires external verification mechanisms. This work provides a measurable framework for inference-layer governability and introduces a taxonomy for evaluating deployment risk in autonomous AI systems.