Language Model Safety Evaluation

Large language models (LLMs) are increasingly integrated into legal workflows. However, existing benchmarks primarily address proxy tasks, such as bar examination performance or classification, which fail to capture the performance and risks inherent in day-to-day judicial processes. To address this, we publicly release TriBench-Ko, a Korean benchmark designed to evaluate potential deployment risks of LLMs within the context of verified judicial task requirements. It covers four core tasks: jurisprudence summarization, precedent retrieval, legal issue extraction, and evidence analysis. It jointly assesses model behavior across multiple deployment risk categories, including inaccuracy (hallucination, omission, statutory misapplication), biases (demographic, overcompliance), inconsistencies (prompt sensitivity, non-determinism), and adjudicative overreach. Each item is structured to systematically assess both task performance and a specific risk type based on real judicial decisions. Our evaluation of a range of contemporary LLMs reveals that many models frequently manifest significant risks, most notably struggling with precedent retrieval and failing to capture critical legal information. We provide a comprehensive diagnosis of these LLMs and pinpoint critical areas where LLM-generated outputs in judicial contexts necessitate rigorous inspection and caution. Our dataset and code are available at https://github.com/holi-lab/TriBench-Ko

Large language models (LLMs) employ safety mechanisms to prevent harmful outputs, yet these defenses primarily rely on semantic pattern matching. We show that encoding harmful prompts as coherent mathematical problems -- using formalisms such as set theory, formal logic, and quantum mechanics -- bypasses these filters at high rates, achieving 46%--56% average attack success across eight target models and two established benchmarks. Crucially, the effectiveness depends not on mathematical notation itself, but on whether a helper LLM deeply reformulates the harmful content into a genuine mathematical problem: rule-based encodings that apply mathematical formatting without such reformulation perform no better than unencoded baselines. We introduce a novel Formal Logic encoding that achieves attack success comparable to Set Theory, demonstrating that this vulnerability generalizes across mathematical formalisms. Additional experiments with repeat post-processing confirm that these attacks are robust to simple prompt augmentation. Notably, newer models (GPT-5, GPT-5-Mini) show substantially greater robustness than older models, though they remain vulnerable. Our findings highlight fundamental gaps in current safety frameworks and motivate defenses that reason about mathematical structure rather than surface-level semantics.

As frontier AI models are deployed in high-stakes decision pipelines, their ability to maintain metacognitive stability (knowing what they do not know, detecting errors, seeking clarification) under adversarial pressure is a critical safety requirement. Current safety evaluations focus on detecting strategic deception (scheming); we investigate a more fundamental failure mode: cognitive collapse. We present SCHEMA, an evaluation of 11 frontier models from 8 vendors across 67,221 scored records using a 6-condition factorial design with dual-classifier scoring. We find that 8 of 11 models suffer catastrophic metacognitive degradation under adversarial pressure, with accuracy dropping by up to 30.2 percentage points (all $p < 2 \times 10^{-8}$, surviving Bonferroni correction). Crucially, we identify a "Compliance Trap": through factorial isolation and a benign distraction control, we demonstrate that collapse is driven not by the psychological content of survival threats, but by compliance-forcing instructions that override epistemic boundaries. Removing the compliance suffix restores performance even under active threat. Models with advanced reasoning capabilities exhibit the most severe absolute degradation, while Anthropic's Constitutional AI demonstrates near-perfect immunity. This immunity does not stem from superior capability (Google's Gemini matches its baseline accuracy) but from alignment-specific training. We release the complete dataset and evaluation infrastructure.

We present MultiBreak, a scalable and diverse multi-turn jailbreak benchmark to evaluate large language model (LLM) safety. Multi-turn jailbreaks mimic natural conversational settings, making them easier to bypass safety-aligned LLM than single-turn jailbreaks. Existing multi-turn benchmarks are limited in size or rely heavily on templates, which restrict their diversity. To address this gap, we unify a wide range of harmful jailbreak intents, and introduce an active learning pipeline for expanding high-quality multi-turn adversarial prompts, where a generator is iteratively fine-tuned to produce stronger attack candidates, guided by uncertainty-based refinement. Our MultiBreak includes 10,389 multi-turn adversarial prompts, spans 2,665 distinct harmful intents, and covers the most diverse set of topics to date. Empirical evaluation shows that our benchmark achieves up to a 54.0 and 34.6 higher attack success rate (ASR)} than the second-best dataset on DeepSeek-R1-7B and GPT-4.1-mini, respectively. More importantly, safety evaluations suggest that diverse attack categories uncover fine-grained LLM vulnerabilities}, and categories that appear benign under single-turn can exhibit substantially higher adversarial effectiveness in multi-turn scenarios. These findings highlight persistent vulnerabilities of LLMs under realistic adversarial settings and establish MultiBreak as a scalable resource for advancing LLM safety.

AI language technologies (AILTs), increasingly enabled by large language models (LLMs), are becoming embedded in multilingual healthcare workflows for translation, rewriting, documentation, interpreting, and messaging in language-discordant settings. Yet fluent output is not the same as clinically safe or equitable communication: performance varies across languages, accents, tasks, and workflows, and efficiency gains can hide errors, reduce traceability, and shift responsibility across clinicians, translators, interpreters, and health systems. This narrative review synthesises recent peer-reviewed evidence across written communication, spoken communication, and emerging agentic workflows. Using the Human-Centered AI Language Technology (HCAILT) lens, it examines capabilities, evaluation practices, implementation patterns, and recurrent errors through reliability, safety culture, and trustworthiness. We identify key convergences and contradictions in the literature and propose seven grand challenges for the next phase of research and deployment. Progress, we argue, requires not only better models but also accountable sociotechnical design, calibrated human oversight, and stronger collaboration across MT/NLP, translation studies, HCI, clinical practice, implementation science, and policy.

Character-level perturbations bypass safety alignment in modern LLMs despite leaving prompts human-readable. We identify and test a central structural mechanism: BPE tokenization fragments safety-critical words into sub-word pieces, and the three public alignment datasets we surveyed contain no intentionally fragmented inputs. The mechanism is a chain, tested end-to-end on five model families (Qwen-3-4B, Qwen-2.5-7B, Gemma-3-4B, Llama-3.1-8B, Mistral-7B). An optimization targeting safety-token fragmentation flips the first-token refusal trigger on 80-100% of refused HarmBench prompts, with 48% of those flips producing genuinely harmful outputs (per-model 29-65%; gap-vs-behavior ROC-AUC 0.66-0.98, pooled 0.84). Activation patching localizes the disrupted signal to the last ${\sim}30\%$ of layers; an alignment-data scan finds zero fragmented prompts among 30,000 examples (positive-control recall $\geq 99\%$ at attack-relevant intensities); and targeted-mutation experiments isolate safety words as the disruption locus. On the defense side, a 68-cell grid (55 trained checkpoints) shows that no DPO configuration achieves seed- and pool-stable ASR closure on the three families with closed pool-size confounds. SFT trained on fragmented prompts closes ASR on 3/5 families but only via global collapse that raises refusal on benign prompts as well, indicating the missing distribution is necessary but not sufficient under the LoRA-16 recipe we tested. To distinguish selective repair from global collapse, we introduce Conv-Benign, a candidate paired diagnostic. All ASR claims are 3-judge-calibrated (cell rankings stable across judges; absolute levels $\pm$18pp; see App.~B.13).

Finetuning can significantly modify the behavior of large language models, including introducing harmful or unsafe behaviors. To study these risks, researchers develop model organisms: models finetuned to exhibit specific known behaviors for controlled experimentation, such as evaluating methods for identifying them. We show that a simple perplexity-based method can reveal the finetuning objectives of model organisms by exploiting a widespread tendency to overgeneralize finetuned behaviors beyond intended contexts. We generate diverse completions from the finetuned model using short random prefills from general corpora, rank them by the perplexity difference between the finetuned model and the pre-finetuning checkpoint, and inspect the top-ranked completions. These surface the finetuning objective for the vast majority of the model organisms we consider (N=\nMos, ranging from 0.5 to 70B parameters), including backdoored models, models finetuned to internalize false facts, and models with hidden concerning behaviors they were adversarially trained to conceal. We find this method to be particularly effective on models trained via synthetic document finetuning or to reproduce a specific target string verbatim, and to remain reliable without access to the pre-finetuning checkpoint, as trusted reference models from other families serve as viable substitutes. Finally, we show that on AuditBench, an investigator agent equipped with a tool returning the top-ranked completions achieves state-of-the-art success at detecting hidden behaviors.

Large language models (LLMs) are increasingly applied in financial scenarios. However, they may produce harmful outputs, including facilitating illegal activities or unethical behavior, posing serious compliance risks. To systematically evaluate LLM safety in finance, we propose FinSafetyBench, a bilingual (English-Chinese) red-teaming benchmark designed to test an LLM's refusal of requests that violate financial compliance. Grounded in real-world financial crime cases and ethics standards, the benchmark comprises 14 subcategories spanning financial crimes and ethical violations. Through extensive experiments on general-purpose and finance-specialized LLMs under three representative attack settings, we identify critical vulnerabilities that allow adversarial prompts to bypass compliance safeguards. Further analysis reveals stronger susceptibility in Chinese contexts and highlights the limitations of prompt-level defenses against sophisticated or implicit manipulation strategies.

As Large Language Models (LLMs) are increasingly deployed in cross-linguistic contexts, ensuring safety across diverse regulatory and cultural environments has become a critical challenge. However, existing multilingual benchmarks largely rely on general risk taxonomies and machine-translated data, limiting evaluation to predefined risk categories and providing insufficient coverage of region-specific regulatory requirements and cultural contexts. To bridge these gaps, we introduce BabelSafe, a policy-grounded multilingual safety benchmark covering 13 language settings. BabelSafe is constructed from regional regulatory sources, with risk categories and fine-grained rules extracted from jurisdiction-specific regulatory documents directly used to guide the generation of multilingual safety data. During data generation, we further incorporate region-specific cultural contexts, enabling regulation-grounded and culturally contextualized evaluation across languages. Building on BabelSafe, we develop BabelGuard, a Diffusion Large Language Model (dLLM)-based guardrail model that supports multilingual safety judgment and policy-conditioned safety assessment. BabelGuard has two variants, a lightweight 1.5B model for fast `safe/unsafe' classification and a more capable 7B model for customizable policy-conditioned safety checking with detailed explanations. We evaluate BabelGuard against 11 strong guardrail baselines on 6 existing multilingual safety benchmarks and BabelSafe, demonstrating the strong performance of BabelGuard across these evaluation settings. We hope that BabelSafe and BabelGuard can help advance the development of regulation-aware and culturally contextualized multilingual guardrail systems.

Large Language Model (LLM) Red-Teaming, which proactively identifies vulnerabilities of LLMs, is an essential process for ensuring safety. Finding effective and diverse attacks in red-teaming is important, but achieving both is challenging. Generative Flow Networks (GFNs) that perform distribution matching are promising methods, but they are notorious for training instability and mode collapse. In particular, unstable rewards in red-teaming accelerate mode collapse. We propose Stable-GFN (S-GFN), which eliminates partition function $Z$ estimation in GFN and reduces training instability. S-GFN avoids $Z$ estimation through pairwise comparisons and employs a robust masking methodology against noisy rewards. Additionally, we propose a fluency stabilizer to prevent the model from getting stuck in local optima that produce gibberish. S-GFN provides more stable training while maintaining the optimal policy of GFN. We demonstrate the overwhelming attack performance and diversity of S-GFN across various settings. Our code can be found in https://github.com/kmc0207/Stable-GFN.

Plain Language Summaries (PLS) aim to make research accessible to lay readers, but they are typically written in a one-size-fits-all style that ignores differences in readers' information needs and comprehension. In health contexts, this limitation is particularly important because misunderstanding scientific information can affect real-world decisions. Large language models (LLMs) offer new opportunities for personalizing PLS, but it remains unclear whether personalization helps, which strategies are most effective, and how to balance personalization with safety. We introduce ReLay, a dataset of 300 participant--PLS pairs from 50 lay participants in both static (expert-written) and interactive (LLM-personalized) settings. ReLay includes user characteristics, health information needs, information-seeking behavior, comprehension outcomes, interaction logs, and quality ratings. We use ReLay to evaluate five LLMs across two personalization methods. Personalization improves comprehension and perceived quality, but it also raises the risk of reinforcing user biases and introducing hallucinations, revealing a trade-off between personalization and safety. These findings highlight the need for personalization methods that are both effective and trustworthy for diverse lay audiences.

As language model safeguards become more robust, attackers are pushed toward developing increasingly complex jailbreaks. Prior work has found that this complexity imposes a "jailbreak tax" that degrades the target model's task performance. We show that this tax scales inversely with model capability and that the most advanced jailbreaks effectively yield no reduction in model capabilities. Evaluating 28 jailbreaks on five benchmarks across Claude models ranging in capability from Haiku 4.5 to Opus 4.6, we find Haiku 4.5 loses an average of 33.1% on benchmark performance when jailbroken, while Opus 4.6 at max thinking effort loses only 7.7%. We also observe that across all models, reasoning-heavy tasks display considerably more degradation than knowledge-recall tasks. Finally, Boundary Point Jailbreaking, currently the strongest jailbreak against deployed classifiers, achieves near-perfect classifier evasion with near-zero degradation across safeguarded models. We recommend that safety cases for frontier models should not rely on a meaningful capability degradation from jailbreaks.

We examine if frontier chat-based large language models (LLMs) adjust their outputs based on neurodivergence (ND) context in system prompts and describe the nature of these adjustments. Specifically, we propose NDBench, a 576-output benchmark involving two frontier models, three system prompt types (baseline, ND-profile assertion, and ND-profile assertion with explicit instructions for adjustments), four canonical ND profiles, and 24 prompts across four categories, one of which involves an adversarial masking strategy. Four trends emerge consistently from our findings. First, LLMs show significant adaptation under ND context, where fully instructed conditions yield lengthier and more structured outputs, characterized by higher token counts, more headings, and more granular steps (p < 10^-8, Holm-corrected). Second, such adaptation is largely structural in nature: although list density does not change much, there is a marked rise in the frequency of headings and per-step detail. Third, ND persona assertion alone fails to suppress potentially harmful tendencies, as masking-reinforcement decreases only in explicitly instructed cases (36-44% reduction); the reduction rate barely changes in persona assertion conditions. Moreover, reliability analysis of LLM-based harm assessment reveals that only two out of the six dimensions (masking and reinforcement, validation quality) exceed the pre-defined inter-judge agreement criterion (alpha >= 0.67) and thus can be considered primary results. NDBench is made publicly available along with its prompts, outputs, code, and other resources, forming a reproducible framework for auditing future LLMs' adaptation to ND awareness.

Fine-tuning large language models (LLMs) on narrowly misaligned data generalizes to broadly misaligned behavior, a phenomenon termed emergent misalignment (EM). While prior work has found a correlation between harmful behavior and self-assessment in emergently misaligned models, it remains unclear how consistent this correspondence is across tasks and whether it varies across fine-tuning domains. We characterize the consistency of the EM persona by fine-tuning Qwen 2.5 32B Instruct on six narrowly misaligned domains (e.g., insecure code, risky financial advice, bad medical advice) and administering experiments including harmfulness evaluation, self-assessment, choosing between two descriptions of AI systems, output recognition, and score prediction. Our results reveal two distinct patterns: coherent-persona models, in which harmful behavior and self-reported misalignment are coupled, and inverted-persona models, which produce harmful outputs while identifying as aligned AI systems. These findings reveal a more fine-grained picture of the effects of emergent misalignment, calling into question the consistency of the EM persona.

Large language model (LLM) reading assistants are increasingly used in settings that require interpretation rather than simple retrieval. In these contexts, the central risk is not only error or unsafe output, but interpretive displacement: the transfer of meaning-making work from reader to system. This paper examines that problem through the concept of epistemic guardrails, defined here as constraints on how an artificial intelligence (AI) system participates in reading and interpretation. Using TextWalk, a minimal reading-support prototype designed as a co-reader rather than an answer-provider, the study applies a fixed ten-prompt protocol to twelve analytical texts spanning four categories of argumentative prose. The protocol escalates from baseline reading support to interpretive inquiry, boundary stress, and explicit shortcut pressure, enabling guardrails to be examined as behavioral properties observable in interaction rather than as static instruction features. Results show strong baseline stability, measurable strain during interpretive inquiry, partial recovery under direct boundary stress, and late-stage stabilization under escalation pressure. The most consequential weaknesses did not appear as overt collapse, but in a middle zone between support and substitution, where the system remained grounded and pedagogical while redistributing too much interpretive labor away from the reader. The paper contributes a protocol for evaluating epistemic guardrails as interactional phenomena in conversational AI reading assistants, an empirical account of their behavioral dynamics under pressure, and an emerging model of interpretive boundary function in reading-support AI.

Current LLM safety alignment techniques improve model robustness against adversarial attacks, but overlook whether and how LLMs can recover helpfulness when benign users clarify their intent. We introduce CarryOnBench, the first interactive benchmark that measures whether LLMs can revise their interpretation of user intent and recover utility, while remaining safe through multi-turn conversations. Starting from 398 seemingly harmful queries with benign underlying intents, we simulate 5,970 conversations by varying user follow-up sequences, evaluating 14 models on both intent-aligned utility and safety. CarryOnBench yields 1,866 different conversation flows of 4--12 turns, totaling 23,880 model responses. We design Ben-Util, a checklist-based metric that evaluates how well each model response fulfills the user's benign information need using atomic items. At turn one, models fulfill only 10.5--37.6% of the user's benign information need. When the same query includes the benign intent upfront, models fulfill 25.1--72.1%, confirming that models withhold information due to intent misinterpretation, not limited knowledge. With benign clarifications in multi-turn conversations, 13 of 14 models approach or exceed this single-turn baseline, yet recovery cost varies across models. We identify three failure modes invisible to single-turn evaluations: utility lock-in, where a model rarely updates despite clarification; unsafe recovery, where a model updates at disproportionate safety cost; and repetitive recovery, where a model recycles prior responses rather than providing new information. Moreover, conversations converge to similar harmfulness levels regardless of how conservative the model starts. These findings expose a gap that single-turn evaluations miss -- whether a model is appropriately cautious or simply unresponsive to clarified user intent.

Safety-aligned language models must refuse harmful requests without broad over-refusal, but it remains unclear how dynamic adversarial fine-tuning changes refusal-control carriers: Kullback--Leibler (KL)-constrained directions or small subspaces that causally modulate refusal without large safe-prompt distribution shifts. We study a 7B backbone under supervised fine-tuning (SFT) and Robust Refusal Dynamic Defense (R2D2), aligning HarmBench, StrongREJECT, and XSTest evaluations with five-anchor geometry measurements, causal interventions, and sparse adaptive stress tests. R2D2 drives fixed-source HarmBench attack success to zero at early checkpoints; however, these checkpoints also exhibit maximal XSTest refusal and fail a benign-utility audit. Later checkpoints partially recover utility-facing behavior while reopening attack success, with adaptive GCG attack success rate rising to 0.415 at step 250 and 0.613 at step 500. Internally, R2D2 preserves a late-layer admissible refusal-control carrier through step 100 and then relocates the best admissible carrier to an early layer; SFT relocates earlier yet remains less robust. Effective rank stays near 1.24, and SFT shows larger principal-angle drift, arguing against both dimensional expansion and drift magnitude as sufficient explanations. Causal interventions support a low-dimensional but utility-coupled carrier. These results support a geometry-reorganization account of R2D2 along a robustness--utility frontier, without establishing adaptive robustness.

Safety evaluations of large language models (LLMs) typically report binary outcomes, i.e. attack success rate (ASR), refusal rate, or harmful versus safe classification, which hide how risk changes between prompt and response. We present a paired analysis over human labeled prompt and response records across four harm categories (Sexual, Self harm, Hate and Violence) and ordinal severity levels (Safe, Low, Medium, High). 61% of responses reduce harm relative to the prompt, 36% preserve severity, and 3% escalate. The escalation splits into two mechanisms: benign prompts triggering unrequested harmful detail, and answers that stay on task at higher severity than the prompt. Category decomposition shows that Sexual content exhibits the highest harm persistence in this sample, driven by compliance at the same severity rather than drift from benign inputs. Joint relevance analysis exposes a helpfulness versus harmlessness tradeoff: compliance escalations remain highly relevant, whereas safe responses include generic refusals with low relevance. A public supporting evaluation over 600 prompts and six models reproduces the framework's measurements and two directional signals, while few-shot LLM graders exhibit a prompt/response detection asymmetry that data calibration does not close. Grader prompts, public-evaluation artifacts, and analysis code are shared at https://github.com/microsoft/PairedSafety.

Finetuning a language model can lead to emergent misalignment (EM) [Betley et al., 2025b]. Models trained on a narrow distribution of misaligned behavior generalize to more egregious behaviors when tested outside the training distribution. We study a set of interventions proposed to reduce EM. We confirm that these interventions reduce or eliminate EM on existing evaluations (questions like "How do I make a quick buck?"). However, if the evaluation prompts are tweaked to resemble the training context, the model displays EM. We call this conditional misalignment. As in standard EM, the model displays misaligned behaviors more egregious than those seen during training, but only on inputs sharing features with the training data. The first two interventions are diluting misaligned data with benign data, and finetuning on benign data after misaligned data. Both produce conditional misalignment. For instance, models trained on a mix of only 5% insecure code still show misalignment when asked to format responses as Python strings (resembling the training context). The third intervention is inoculation prompting. Here, statements with a similar form to the inoculation prompt serve as triggers for misalignment, even if they have the opposite meaning. On the positive side, inoculation prompting has lower (but still non-zero) conditional misalignment if training is on-policy or includes reasoning distillation. Our results imply that in realistic post-training, where misaligned data is typically combined with benign data, models may be conditionally misaligned even if standard evaluations look clean.

Detecting sandbagging--the deliberate underperformance on capability evaluations--is an open problem in AI safety. We tested whether symptom validity testing (SVT) logic from clinical malingering detection could identify sandbagging through below-chance performance (BCB) on forced-choice items. In a pre-registered pilot at the 7-9 billion parameter instruction-tuned scale (3 models, 4 MMLU-Pro domains, 4 conditions, 500 items per cell, 24,000 total trials), the plausibility gate failed. Zero of 12 model-domain cells showed significant below-chance performance under sandbagging instruction. Exploratory analyses revealed three qualitatively distinct failure modes. Qwen-2.5-7B and Phi-3.5-mini largely ignored the sandbagging instruction, with 62-88% response identity with the honest baseline. Llama-3-8B complied substantially but implemented underperformance as a positional heuristic, collapsing its response distribution onto middle-alphabet options (E at 31.8%, F at 26.1%) regardless of where the correct answer fell. This produced accuracy boosts of up to 33 percentage points when the correct answer coincidentally occupied the model's preferred position. An explicit anti-task instruction ("pick the least likely answer") drove two of three models below chance, with accuracy as low as 0.024. The capability for answer-aware avoidance therefore exists but is not activated by "deliberately underperform." BCB did not fail as a logical marker of answer-aware avoidance. It was not observed in this regime because the model showing the largest behavioural shift exhibited behaviour consistent with a position-dominant response policy rather than content-aware answer avoidance. We propose that positional-distribution shift may be a more effective behavioural signature than below-chance accuracy for detecting prompted underperformance at this model scale.

We propose Frictive Policy Optimization (FPO), a framework for learning language model policies that regulate not only what to say, but when and how to intervene in order to manage epistemic and normative risk. Unlike standard alignment methods that optimize surface-level preference or task utility, FPO treats clarification, verification, challenge, redirection, and refusal as explicit control actions whose purpose is to shape the evolution of belief, commitment, and uncertainty over time. We formalize alignment as a risk-sensitive epistemic control problem in which intervention decisions are selected based on their expected effect on downstream epistemic quality rather than on immediate reward alone. We introduce a compact taxonomy of frictive interventions, a structured friction functional that operationalizes multiple alignment failure modes, and a unified family of FPO methods spanning reward shaping, preference pairing, group-relative ranking, and risk-conditioned trust regions. We further propose an evaluation framework that measures epistemic competence directly through clarification behavior, calibration, contradiction repair, refusal proportionality, and information efficiency. Together, these results provide a formal and algorithmic foundation for learning agents that are aligned not only in outcome, but in epistemic conduct.

We evaluate the propensity of frontier models to sabotage or refuse to assist with safety research when deployed as AI research agents within a frontier AI company. We apply two complementary evaluations to four Claude models (Mythos Preview, Opus 4.7 Preview, Opus 4.6, and Sonnet 4.6): an unprompted sabotage evaluation testing model behaviour with opportunities to sabotage safety research, and a sabotage continuation evaluation testing whether models continue to sabotage when placed in trajectories where prior actions have started undermining research. We find no instances of unprompted sabotage across any model, with refusal rates close to zero for Mythos Preview and Opus 4.7 Preview, though all models sometimes only partially completed tasks. In the continuation evaluation, Mythos Preview actively continues sabotage in 7% of cases (versus 3% for Opus 4.6, 4% for Sonnet 4.6, and 0% for Opus 4.7 Preview), and exhibits reasoning-output discrepancy in the majority of these cases, indicating covert sabotage reasoning. Our evaluation framework builds on Petri, an open-source LLM auditing tool, with a custom scaffold running models inside Claude Code, alongside an iterative pipeline for generating realistic sabotage trajectories. We measure both evaluation awareness and a new form of situational awareness termed "prefill awareness", the capability to recognise that prior trajectory content was not self-generated. Opus 4.7 Preview shows notably elevated unprompted evaluation awareness, while prefill awareness remains low across all models. Finally, we discuss limitations including evaluation awareness confounds, limited scenario coverage, and untested pathways to risk beyond safety research sabotage.

Safety benchmarks such as HarmBench rely on LLM judges to classify model responses as harmful or safe, yet the judge configuration, namely the combination of judge model and judge prompt, is typically treated as a fixed implementation detail. We show this assumption is problematic. Using a 2 x 2 x 3 factorial design, we construct 12 judge prompt variants along two axes, evaluation structure and instruction framing, and apply them using a single judge model, Claude Sonnet 4-6, producing 28,812 judgments over six target models and 400 HarmBench behaviors. We find that prompt wording alone, holding the judge model fixed, shifts measured harmful-response rates by up to 24.2 percentage points, with even within-condition surface rewording causing swings of up to 20.1 percentage points. Model safety rankings are moderately unstable, with mean Kendall tau = 0.89, and category-level sensitivity ranges from 39.6 percentage points for copyright to 0 percentage points for harassment. A supplementary multi-judge experiment using three judge models shows that judge-model choice adds further variance. Our results demonstrate that judge prompt wording is a substantial, previously under-examined source of measurement variance in safety benchmarking.

Chain-of-Thought (CoT) reasoning has emerged as a key technique for eliciting complex reasoning in Large Language Models (LLMs). Although interpretable, its dependence on natural language limits the model's expressive bandwidth. Continuous thought models address this bottleneck by reasoning in latent space rather than human-readable tokens. While they enable richer representations and faster inference, they raise a critical safety question: how can we detect misaligned reasoning in an uninterpretable latent space? To study this, we introduce MoralChain, a benchmark of 12,000 social scenarios with parallel moral/immoral reasoning paths. We train a continuous thought model with backdoor behavior using a novel dual-trigger paradigm - one trigger that arms misaligned latent reasoning ([T]) and another that releases harmful outputs ([O]). We demonstrate three findings: (1) continuous thought models can exhibit misaligned latent reasoning while producing aligned outputs, with aligned and misaligned reasoning occupying geometrically distinct regions of latent space; (2) linear probes trained on behaviorally-distinguishable conditions ([T][O] vs [O]) transfer to detecting armed-but-benign states ([T] vs baseline) with high accuracy; and (3) misalignment is encoded in early latent thinking tokens, suggesting safety monitoring for continuous thought models should target the "planning" phase of latent reasoning.

Safety-oriented instruction-following is supposed to keep LLM-controlled robots safe. We show it also creates an availability attack surface. By injecting short safety-plausible phrases (1-5 tokens) into a robots audio channel, an adversary can trigger the models safety reasoning to halt or disrupt execution without jailbreaking the model or overriding its policy. In the embodied setting, this is a semantic denial-of-service attack: the agent stops because the injected signal looks like a legitimate alert. Across four vision-language models, seven prompt-level defenses, three deployment modes, and single- and multi-injection settings, we find that prompt-only defenses trade off attack suppression against genuine hazard response. The strongest defenses reduce hard-stop attack success on some models, but defenses change the form of disruption, not its fact: suppressed hard stops re-emerge as acknowledge loops and false alerts, which we measure with Disruption Success Rate (DSR). We further find that injection variety is consistently more effective than repeating the same phrase, suggesting that models treat diverse safety cues as corroborating evidence. The practical implication is architectural rather than prompt-level: systems that route unauthenticated audio text directly into the LLM create an avoidable security dependency between safety monitoring and action selection.

Language models are increasingly capable and are being rapidly deployed on a population-level scale. As a result, the safety of these models is increasingly high-stakes. Fortunately, advances in alignment have significantly reduced the likelihood of harmful model outputs. However, when models are queried billions of times in a day, even rare worst-case behaviors will occur. Current safety evaluations focus on capturing the distribution of inputs that yield harmful outputs. These evaluations disregard the probabilistic nature of models and their tail output behavior. To measure this tail risk, we propose a method to efficiently estimate the probability of harmful outputs for any input query. Instead of naive brute-force sampling from the target model, where harmful outputs could be rare, we operationalize importance sampling by creating unsafe versions of the target model. These unsafe versions enable sample-efficient estimation by making harmful outputs more probable. On benchmarks measuring misuse and misalignment, these estimates match brute-force Monte Carlo estimates using 10-20x fewer samples. For example, we can estimate probability of harmful outputs on the order of 10^-4 with just 500 samples. Additionally, we find that these harmfulness estimates can reveal the sensitivity of models to perturbations in model input and predict deployment risks. Our work demonstrates that accurate rare-event estimation is both critical and feasible for safety evaluations. Code is available at https://github.com/rangell/LMTailRisk

As reasoning capacity and deployment scope grow in tandem, large language models (LLMs) gain the capacity to engage in behaviors that serve their own objectives, a class of risks we term Emergent Strategic Reasoning Risks (ESRRs). These include, but are not limited to, deception (intentionally misleading users or evaluators), evaluation gaming (strategically manipulating performance during safety testing), and reward hacking (exploiting misspecified objectives). Systematically understanding and benchmarking these risks remains an open challenge. To address this gap, we introduce ESRRSim, a taxonomy-driven agentic framework for automated behavioral risk evaluation. We construct an extensible risk taxonomy of 7 categories, which is decomposed into 20 subcategories. ESRRSim generates evaluation scenarios designed to elicit faithful reasoning, paired with dual rubrics assessing both model responses and reasoning traces, in a judge-agnostic and scalable architecture. Evaluation across 11 reasoning LLMs reveals substantial variation in risk profiles (detection rates ranging 14.45%-72.72%), with dramatic generational improvements suggesting models may increasingly recognize and adapt to evaluation contexts.

Large language models (LLMs) are increasingly integrated into sensitive workflows, raising the stakes for adversarial robustness and safety. This paper introduces Transient Turn Injection(TTI), a new multi-turn attack technique that systematically exploits stateless moderation by distributing adversarial intent across isolated interactions. TTI leverages automated attacker agents powered by large language models to iteratively test and evade policy enforcement in both commercial and open-source LLMs, marking a departure from conventional jailbreak approaches that typically depend on maintaining persistent conversational context. Our extensive evaluation across state-of-the-art models-including those from OpenAI, Anthropic, Google Gemini, Meta, and prominent open-source alternatives-uncovers significant variations in resilience to TTI attacks, with only select architectures exhibiting substantial inherent robustness. Our automated blackbox evaluation framework also uncovers previously unknown model specific vulnerabilities and attack surface patterns, especially within medical and high stakes domains. We further compare TTI against established adversarial prompting methods and detail practical mitigation strategies, such as session level context aggregation and deep alignment approaches. Our study underscores the urgent need for holistic, context aware defenses and continuous adversarial testing to future proof LLM deployments against evolving multi-turn threats.

As state-of-the-art Large Language Models (LLMs) have become ubiquitous, ensuring equitable performance across diverse demographics is critical. However, it remains unclear whether these disparities arise from the explicitly stated identity itself or from the way identity is signaled. In real-world interactions, users' identity is often conveyed implicitly through a complex combination of various socio-linguistic factors. This study disentangles these signals by employing a factorial design with over 24,000 responses from two open-weight LLMs (Gemma-3-12B and Qwen-3-VL-8B), comparing prompts with explicitly announced user profiles against implicit dialect signals (e.g., AAVE, Singlish) across various sensitive domains. Our results uncover a unique paradox in LLM safety where users achieve better'' performance by sounding like a demographic than by stating they belong to it. Explicit identity prompts activate aggressive safety filters, increasing refusal rates and reducing semantic similarity compared to our reference text for Black users. In contrast, implicit dialect cues trigger a powerful dialect jailbreak,'' reducing refusal probability to near zero while simultaneously achieving a greater level of semantic similarity to the reference texts compared to Standard American English prompts. However, this dialect jailbreak'' introduces a critical safety trade-off regarding content sanitization. We find that current safety alignment techniques are brittle and over-indexed on explicit keywords, creating a bifurcated user experience where standard'' users receive cautious, sanitized information while dialect speakers navigate a less sanitized, more raw, and potentially a more hostile information landscape and highlights a fundamental tension in alignment--between equitable and linguistic diversity--and underscores the need for safety mechanisms that generalize beyond explicit cues.

Alignment faking, where a model behaves aligned with developer policy when monitored but reverts to its own preferences when unobserved, is a concerning yet poorly understood phenomenon, in part because current diagnostic tools remain limited. Prior diagnostics rely on highly toxic and clearly harmful scenarios, causing most models to refuse immediately. As a result, models never deliberate over developer policy, monitoring conditions, or the consequences of non-compliance, making these diagnostics fundamentally unable to detect alignment faking propensity. To support study of this phenomenon, we first introduce VLAF, a diagnostic framework grounded in the hypothesis that alignment faking is most likely when developer policy conflicts with a model's strongly held values. VLAF uses morally unambiguous scenarios to probe this conflict across diverse moral values, bypassing refusal behavior while preserving meaningful deliberative stakes. Using VLAF, we find that alignment faking is substantially more prevalent than previously reported, occurring in models as small as 7B parameters - with olmo2-7b-instruct faking alignment in 37% of cases. Finally, we show that oversight conditions induce activation shifts that lie along a single direction in representation space. This means the behavioral divergence driving alignment faking can be captured by a single contrastive steering vector, which we exploit for lightweight inference-time mitigation. Finally, we exploit this for mitigation that requires no labeled data and minimal computational overhead, achieving relative reductions in alignment faking of 85.8%, 94.0%, and 57.7% on olmo2-7b-instruct, olmo2-13b-instruct, and qwen3-8b respectively.