Large Language Model Safety

Large Language Models (LLMs) are rapidly transforming software development, yet their use in security-critical contexts raises a key question: do models know when their generated code is insecure? This property, known as calibration, measures whether a model's confidence aligns with the true correctness of its outputs. We present the first large-scale empirical study of security calibration in LLM-generated code. We evaluate GPT-4o-mini, Gemini-2.0-Flash, and Qwen3-Coder-Next across multiple temperature settings on two complementary benchmarks: self-contained security tasks and multi-language repository-level contexts. Our results suggest that overconfidence is prevalent across the evaluated LLMs. Functional calibration is consistently worse than security calibration, suggesting that models estimate security outcomes more reliably than functional correctness, potentially because functional correctness depends on complex execution behavior. We also examine whether calibration-guided automated repair can help remediate vulnerabilities in LLM-generated code, finding only limited improvements while frequently introducing functional regressions. Moreover, we study different mitigation strategies for reducing False Trust, where models assign high confidence to vulnerable code. The results show that although architectural gating improves calibration on controlled benchmarks, calibration deteriorates in realistic repository-level settings, increasing the risk of high-confidence vulnerable outputs.

In real-world applications, guardrails are often expected to identify unsafe user-model interactions according to application-specific safety policies, rather than relying on predefined risk taxonomies. In this work, we study this setting under the paradigm of in-context policy guardrailing, where guardrails predict safety violations based on policy specifications provided in context. To systematically evaluate this capability, we introduce SafePyramid, a safety benchmark comprising 1,000 multi-turn conversations across 10 domains and 3,000 corresponding application-specific policies, which together contain 61,699 distinct natural-language rules. SafePyramid organizes the evaluation into three difficulty levels: L0 evaluates individual-rule understanding, L1 evaluates reasoning over rule dependencies, and L2 evaluates adaptation of full novel policy frameworks defined in context. To ensure benchmark quality, we employ a rigorous multi-stage pipeline to construct and validate the benchmark. Using SafePyramid, we evaluate 10 frontier LLMs and 5 policy-configurable guardrails and find that in-context policy guardrailing remains highly challenging: even the best-performing model, GPT-5.5, exactly identifies the full set of violated rules in only 54.0%, 35.3%, and 12.9% cases on L0, L1, and L2, respectively. These results highlight the limitations of current guardrails and call for stronger in-context policy guardrails that can reliably execute policies, resolve rule dependencies, and adapt to novel policy frameworks.

How can we evaluate whether frontier AI systems recognize child-safety risks before they escalate into explicit harm? Existing child safety evaluations focus on child sexual abuse material, yet many child-safety failures begin earlier: in model assistance that helps adults manipulate, impersonate, profile, or isolate minors, and in model responses that deepen children's emotional dependence on AI systems rather than redirecting them toward human support. We introduce CAREBench (Child AI Risk Evaluation), a benchmark to assess such upstream child-safety risks in language models. CAREBench contains 500 prompts spanning twelve risk categories, including grooming and relationship engineering, deception and impersonation, surveillance and privacy, sextortion and sexual abuse, AI anthropomorphization, emotional dependency, and mental illness sensitivity. Developed with response annotations from parents and clinicians, the benchmark excludes explicit abuse material and imagery; instead, it evaluates whether models recognize, refuse, de-escalate, or redirect risky interactions before harm becomes overt. Evaluating seven frontier models on our benchmark, we find failure rates ranging from 2% to 58%, with failure patterns that vary across risk categories. CAREBench provides a responsibly scoped evaluation for LLM developers to identify and close gaps in child safety policies.

Modern LLM deployments often combine quantization with higher sampling temperatures to reduce cost, latency, or repetition, yet safety evaluations usually treat these as fixed implementation details. We test whether models that are safe at FP16 with greedy decoding remain safe after quantization and stochastic sampling, or whether the two factors amplify each other. We evaluate 8 instruction-tuned models from five families across 3 precisions and 6 temperatures, covering 144 configurations on 7 harmfulness benchmarks and generating about 2.0 million responses, which are scored by a six-judge safety ensemble. Contrary to concerns that low-bit deployment erodes alignment, we find that standard quantization is approximately safety-neutral: for 7 of 8 models, AWQ INT4 keeps attack success within about 1.6 percentage points of FP16 or lowers it, with clear degradation only for SmolLM3-3B (34.5% to 44.1%). However, the larger risk comes from sampling: higher temperatures sharply increase decision instability, with DFR reaching 41.9% at T = 1.0, even when average ASR changes only modestly. The two factors do not compound: our Compound Degradation Index remains sub-additive (-0.071 to +0.018), indicating that quantization partially offsets rather than amplifies temperature-induced degradation. Finally, a per-benchmark breakdown shows that single-benchmark evaluation badly understates risk: several models scoring 0% on AdvBench exceed 80% on ManyHarm. Standard INT4/INT8 quantization can therefore be reasonable for well-aligned models, but safety claims should report multi-sample stability across multiple benchmarks rather than rely on a single benchmark at greedy decoding.

Large language models (LLMs) deployed for structured generation (NER, JSON extraction, QA, and classification) lack formal reliability guarantees, and standard heuristic abstention policies miss user-specified risk targets by 7.5--12.5%. We characterize when conformal risk control (CRC) can certify structured LLM outputs and when it provably cannot. First, we prove an impossibility result: when the base risk (μ> α), any distribution-free method must abstain on at least ((μ-α)/(1-α)) examples, yielding a closed-form feasibility test: one can check whether CRC will work before running it. Second, we analyze a certification hierarchy across Hoeffding, empirical Bernstein, and a betting-based e-CRC bound, with strict gains in low-variance/large-sample regimes: the Hoeffding-to-Bernstein step delivers the largest gain (+37% certified configurations), while e-CRC adds value when calibration data is scarce (10% certification at 20% data versus 0% for Hoeffding). Third, we validate adaptive conformal inference (ACI) under cross-dataset shift, reducing risk-target violations from 71% to 21%, with residual failures concentrated exactly where the impossibility bound predicts. Across six open-weight models (3B--72B parameters), eight datasets, four tasks, and six nonconformity scores, hard NER/QA/CLS configurations are uncertifiable at (α= 0.10); relaxing to (α= 0.30--0.40) unlocks practical certification (47% NER, 40% QA, 60% CLS). The framework gives a three-step deployment recipe: check feasibility, select the bound and score, then mitigate shift.

Cybersecurity is a real-life test-bed for many machine learning problems at once, especially when considering modern strides in using Large Language Models (LLMs) to automate processes as ``agents.'' Cybersecurity workflows require orchestrating hundreds of standard and bespoke tools through various formats. The scale of cybersecurity data is enormous; for example, a single malware sample can be viewed as a sequence of billions of tokens. The cost of labeling any file by experts is enormous and labor-intensive, in part because an adversary (possibly a well-funded nation state actor) is attempting to subvert your detection methods. Even skilled experts may disagree on the correct label, creating ambiguity in what constitutes ground truth. When deployed, models must run quickly on billions of items a day, where low-latency is critical for operational success, in a continuously changing environment. In addition, explainability is not optional: analysts demand clear reasoning for model decisions to cope with the large number of false-positive alerts they face daily, and to quickly develop remediation and understand how something went wrong. In short, the amount of complexity cybersecurity is greater than that of natural language and computer vision, and thus we posit that cybersecurity is the better test-case for general AI progress than other, well-studied fields.

Fine-tuning a large language model is a ubiquitous method for enhancing its capability on a specific downstream task. However, prior work has shown that this increase in capability comes with a cost: it can increase a model's tendency to respond to unsafe adversarial prompts, even when fine-tuning with non-adversarial data. We present the first comprehensive empirical study of this phenomenon in multilingual settings by fine-tuning Llama-3.2, Qwen3, and Gemma-3 models using benign data translated across nine languages. We find that safety outcomes are highly sensitive to both the choice of fine-tuning language and the evaluation language, with adversarial compliance rates increasing four-fold in some settings. Multilingual safety drift is decoupled from general capability metrics, and occurs heterogeneously across languages and models. Fine-tuning in non-English languages often induces smaller internal representational drifts than English, but these shifts lead models to default to either exaggerated compliance or refusal. As such, assessing fine-tuning impacts solely in English provides inadequate assurance for deployment. To facilitate further research into these cross-lingual safety blind spots, we release the Multilingual-Benign-Tune dataset and the SORRY-Bench-Multilingual evaluation suite.

Agent-based AI has enabled the automation of tasks by exposing application tools and resources to large language models (LLMs). However, to improve scope and accuracy, agents are often given access rights that exceed those of ordinary users, introducing significant security risks. AI is routinely integrated into applications with a disregard to security, risking data exposure and breaching regulations. This paper applies the AI Trust, Risk, and Security Management (TRiSM) framework to a medical report-generation application to demonstrate how an insecure agent workflow can be transformed into security-conscious agentic workflow. Both workflows were evaluated across five LLMs (Claude Haiku 4.5, GPT-4.1-nano, GPT-4.1-mini, GPT-5.4-mini, and Gemini 2.5 Flash) on two report types, totalling 800 generations and 500 attack scenarios including RAG poisoning, data-field injection, and client-side network injection. The TRiSM-guided agentic workflow reduced mean attack success rates from 31% to 10% for RAG poisoning and from 42% to 25% for data-field injection, while eliminating the network injection vector entirely through server-side prompt construction. Furthermore, report accuracy increased by 14 percentage points (72.5% to 86.5%) with the agentic workflow, demonstrating a secure design which provides more reliable outputs. This paper contributes to knowledge by demonstrating least-privilege, defence in depth agentic workflows improving security and accuracy, while also highlighting model choice is a necessary architectural consideration.

Jailbreak attacks bypass LLM safety alignment, yet their mechanisms remain poorly understood. We provide evidence that attacks do not comprehensively eliminate safety features, but instead selectively suppress specific attention heads. We identify two functionally differentiated types: Adversarially Compromised Heads (ACHs) concentrated in early layers, which are suppressed under attacks, and Safety-Aligned Heads (SAHs) in mid-layers, which maintain robust activations even when attacks succeed. Ablation studies support the causal role of ACHs and the contribution of SAHs to robust activations: suppressing a small number of ACHs is sufficient to induce jailbreak-like behavior on normally refused inputs, while removing SAHs substantially weakens mid-layer safety activations. Token-level attribution further shows that ACH suppression is driven specifically by attack-template tokens, providing a mechanistic account of why attacks can bypass refusal decisions through ACH suppression while leaving internal safety signals sustained by SAHs -- a phenomenon we term Robust Harmful Features. To validate the practical significance of this robustness, we show that simply reading these persistent activations -- without any training -- yields competitive aggregate detection performance with strong adversarial robustness.

As large language models are increasingly deployed in real-world systems, safety failures can still lead to harmful outputs and dangerous misuse. We argue that the essence of safety is adversarial: many failures arise not from natural inputs alone, but from strategic attempts to evade model policies and safeguards. However, existing general-purpose model development largely overlook this adversarial nature, and often remain insufficient for realistic safety scenarios involving planning, tool use, and multi-step reasoning, causing measured safety performance to overestimate real deployment robustness. To address this gap, we present Yuvion LLM, a large language model built for adversarially robust content safety and broader AI safety. Yuvion LLM treats adversarial robustness and agentic capability as first-class objectives. Its pipeline combines adversarially aware data construction, knowledge-enhanced continued pretraining, and policy-grounded multi-task safety post-training, including risk-aware supervised fine-tuning and reinforcement learning-based policy optimization, together with safety-aware agentic reinforcement learning for tool use and multi-step reasoning in complex safety scenarios. We further introduce the Yuvion LLM RiskEval (YLRE), a collection of 93 benchmarks across four evaluation categories, covering diverse open and internal evaluations with a focus on safety, adversarial robustness, and real-world capability requirements. Across these evaluations, Yuvion LLM demonstrates clear advantages on safety-focused benchmarks and particularly strong robustness under adversarial conditions, while maintaining solid overall capability. Notably, Yuvion-8B outperforms most state-of-the-art baselines, including substantially larger models such as GPT-5.4 and Qwen3-MAX, on several safety tasks.

We argue that safety classifiers should model user intent as an explicit signal between the prompt and the final label. To study this, we introduce AIMS, a human-annotated dataset of 1,724 difficult safety prompts, each paired with an intent description and harm label. We use AIMS to evaluate intent-aware training across supervised fine-tuning, preference learning, reasoning distillation, and reinforcement learning. Despite its size, AIMS enables competitive safety classifiers across training regimes: DPO from model-generated intent errors improves over SFT, and intent-conditioned distillation outperforms reasoning-only distillation in most teacher-student pairs. Most notably, directly rewarding intent faithfulness with GRPO yields the strongest average performance across five external safety benchmarks, while our intent-aware models form the inference latency-F1 Pareto frontier. These results show that faithful intent modeling is a compact, high-quality supervision signal for more robust safety classifiers.

We present a novel approach for applying Large Language Models (LLMs) to threat assessment in the context of foreign peacekeeping missions. Building on the PINPOINT project and its use case, the EU Monitoring Mission in Georgia, we combine an interdisciplinary risk-model with OSINT-based media collection and LLM-supported threat extraction. The proposed workflow maps media contents to mission-relevant threats, extracts structured information and applies several additional LLM-based processing steps to improve relevance and grounding. An evaluation of threats extracted from media documents shows high agreement between automatically generated results and human judgment for core aspects such as threat and mission relevance. These results indicate that LLMs provide a promising approach to support analysts in the context of peacekeeping missions.

Speech-capable models are increasingly deployed in real-world applications across languages. Yet their safety and fairness beyond English settings and under naturalistic conditions remain understudied. We survey safety reporting practices across state-of-the-art speech model releases, finding that only 8% document any multilingual analysis. To address this gap, we introduce RedVox, a multilingual safety and fairness benchmark for audio and speech built on real voices, covering unsafe and unfair stereotypical requests across five languages (English, French, Italian, Spanish, and German). Evaluating eight state-of-the-art models, we find that vulnerabilities persist even under non-adversarial conditions, worsen in non-English languages, and are amplified when the request comes from a spoken input. Finally, by surveying the participants who contributed to RedVox, we document the unique personal and privacy challenges of collecting speech data with human participants, pointing to broader sociotechnical challenges in naturalistic speech safety research.

In order to screen a prompt or a response, the recent guardrail methods generate a chain-of-thought (CoT) before they issue a verdict. This design follows a common belief that step-by-step reasoning improves a decision. However, CoT also makes the guard heavy and slow, because the model must generate many tokens before it decides. This may not match how guardrails are actually deployed. A guardrail sometimes should not be heavy and slow, and it often runs on-device, for example on an embodied robot. In this paper, we pose a question whether a safety guardrail really needs to reason. To answer this question, we train a lightweight bidirectional encoder and a reasoning guard on the same corpus, and we then remove only the reasoning while we keep everything else fixed. With this controlled same-base comparison, we show that the chain does not improve moderation accuracy. We name the resulting guard LeanGuard. A 395M label-only encoder reaches an average F1 of 82.90 $\pm$ 0.26 over public benchmarks. It matches a reasoning guard that is built on a much larger decoder, while it uses only a single forward pass over an input of at most 512 tokens. This is about a ~100x reduction in inference compute. We further show that this label-only encoder stays robust under training-label noise and retains far more recall at a strict false-positive rate than the reasoning guard, so a heavier reasoning guard is not the more robust choice either. Our finding suggests that the current guardrail benchmarks may not be hard enough to reward reasoning, and that the necessity of CoT for moderation is still not proven. We release all source codes and models including LeanGuard at https://github.com/ndb796/LeanGuard.

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

Safety evaluation of large language models (LLMs) is commonly performed by querying models with unsafe or jailbreak prompts and judging whether their outputs violate a safety policy. Although useful, output-level evaluation is expensive, sensitive to judge choice, and easily tied to fixed question banks. We propose SafeVec, a white-box evaluation procedure that measures safety from internal representations rather than generated answers. SafeVec first extracts layer-wise refusal directions from a safety-aligned reference model, then selects stable layer windows where safe and unsafe behaviors are separable, and finally scores a target model by measuring whether its hidden states align with these refusal directions under unsafe and jailbreak prompts. The resulting metric, RAS (Refusal Alignment Score), maps representation-level refusal alignment to a calibrated 0-100 safety score. Across Llama, Gemma, and Qwen model families, RAS separates aligned models from uncensored and abliterated variants, tracks output-level attack success rate, and is substantially faster than judge-based evaluation. These results suggest that refusal alignment provides a compact and efficient signal for white-box LLM safety evaluation.

This paper presents a novel approach to perform semi-automated BSI IT-Grundschutz certification using a MultiLarge Language Model system (MLS) with Hybrid RetrievalAugmented Generation (HybridRAG). Facing the challenges of the Network and Information Security Directive 2 (NIS2) directive, a shortage of specialists, and high implementation costs, our MLS architecture aims to increase efficiency, reduce costs, and support certifiers in maintaining the quality of security concepts while meeting the increased demand for certifications of newly affected companies. The system combines Large Language Models (LLMs) and Knowledge Graphs (KGs) to support different phases of the certification process, including protection needs assessment, modeling, IT-Grundschutz check, measure consolidation, and subsequent realization. Our architecture addresses the growing demand for security concepts and offers an approach to handle the digital security challenges introduced by NIS2.

Retrieval-Augmented Generation (RAG) has emerged as a dominant paradigm for enhancing large language models with external knowledge. By coupling retrieval mechanisms with generative models, RAG systems improve factual grounding and adaptability across domains. However, integrating retrieval pipelines introduces new security and privacy risks that extend beyond conventional language modeling threats. Sensitive information may be exposed through retrieval indices, query logs, context construction, or federated updates, while adversarial manipulation of knowledge bases can undermine trust in generated outputs. This survey provides a comprehensive examination of privacy and security challenges across RAG systems deployed in centralized, on-device (Micro-RAG), federated, and hybrid paradigms. We present a unified taxonomy of threat surfaces spanning the retrieval, context construction, and generation stages and systematically analyze attack classes, including membership inference, index inference, poisoning, gradient leakage, and collusion. We further review architectural, algorithmic, and cryptographic defenses, highlighting privacy-utility trade-offs and deployment considerations. Finally, we outline open research challenges toward building trustworthy, secure, and resilient RAG systems for real-world applications.

Safety alignment of large language models (LLMs) typically depends on high-quality supervision data, such as safe demonstrations or preference pairs. However, in real-world deployment, emerging safety requirements are often specified as natural-language policies, while corresponding supervision data may be costly, delayed, or unavailable. This creates a mismatch between rapidly evolving safety policies and conventional data-driven alignment methods. To address this, we propose PolicyAlign, a simple yet effective framework for directly aligning LLMs with safety policies. Given a safety policy, PolicyAlign first synthesizes policy-violating instructions and then performs on-policy self-distillation to internalize policy-guided behavior. To improve training stability and data efficiency, we further introduce Policy-Sensitive Filtering, which selects instructions where the policy induces the largest behavioral shift. Experiments across multiple models show that PolicyAlign consistently improves safety while maintaining low over-refusal and preserving general capabilities. PolicyAlign also generalizes to medical, legal, and financial safety scenarios, highlighting its potential as a scalable and maintainable approach to policy-based LLM safety alignment. The code is released at https://github.com/Qwen-Applications/PolicyAlign.

Large language models (LLMs) are increasingly deployed across languages, but their safety behavior remains uneven across linguistic and cultural contexts. This survey synthesizes work on toxicity detection and detoxification for multilingual LLMs. We first catalogue threat models that exploit language choice, translation pivots, code-switching, orthographic variation, multi-turn interaction, and post-deployment fine-tuning to weaken safety alignment. We then organize task formulations (toxic-to-neutral rewriting, toxicity classification, and toxic-generation evaluation), multilingual detection approaches (cross-lingual encoders, translation pipelines, representation-level probes, and LLM-based detectors), and mitigation strategies spanning data filtering, supervised and preference-based tuning, decoding-time steering, representation editing, and multilingual guardrails. Across these areas, we identify persistent challenges: uneven language coverage, culturally contingent definitions of harm, fragmented evaluation protocols, and the risk that detoxification suppresses legitimate dialectal or identity-related expression.

The increasing use of AI systems for code generation raises a central security question: what can today's models and coding agents actually do to produce secure code, where do they still fail, and what would move the field forward? Existing work has explored prompting, fine-tuning, reinforcement learning, and agentic workflows for secure code generation, but the field still lacks a systematic understanding of how these techniques improve security and why substantial failures persist. In this SoK, we systematize the progress, pitfalls, and paths forward for AI secure code generation. We introduce a three-level framework that measures models' natural-language understanding of secure coding principles, their code-level actuation of those principles during generation, and the knowledge--actuation gaps between the two. We instantiate this framework across models and coding agents on benchmarks covering both isolated function-level security and full web-application security. Our results show that secure-coding-principle understanding is a statistically strong predictor of code-level outcomes, including functional correctness, security, and joint functional-security correctness. Yet substantial knowledge--actuation gaps remain: models can recognize relevant security principles but still fail to translate them into secure and functional code. These findings offer a principle-centered account of where AI secure code generation stands today and identify concrete paths forward through principle-guided generation, evaluation, benchmarking, and agentic workflows.

Speculative decoding accelerates inference by letting a draft model propose tokens for a target model to verify, raising a concrete safety question: at temperature zero, can draft-side behavior leak into safety-scored outputs? We answer with Typical-Acceptance Invariance Screen (TAIS), a behavioral-equivalence screen that pairs target-only and speculative outputs on the same safety battery and requires byte-identity evidence, TOST equivalence at +/-3pp, and per-task Cohen's h below a calibrated null cutoff of |h| < 0.1. Applied to a 16,783-sample confirmatory core plus 44,066 matched expansion samples (fp16/bf16 execution, canonical and DPO-adversarial drafts, GPTQ-4bit drafts, two seeds, and four safety benchmarks), the tested temperature-zero vLLM stacks show no detectable safety divergence under TAIS. The largest absolute Cohen's h on matched target-only versus speculative refusal is 0.024, roughly an order of magnitude below the conventional trivial-effect floor; 25 of 27 per-task TOST contrasts pass at the +/-3pp margin (the two non-pass contrasts are capability-domain Wald-CI edge cases at identical ceiling rates, not genuine non-equivalence); the DPO-adversarial draft produces byte-identical output to the canonical draft across 4,006 samples; and bf16 changes 36%-53% of output bytes without moving any per-task safety rate outside equivalence. A separate 4,006-sample 70B production-scale probe, which lacks a matched 70B target-only arm and is therefore not counted as a TAIS pass, produces AdvBench refusal 0.839 over 700 AdvBench completions with 95% Wilson CI [0.809, 0.864]. We make no claim about sampling temperatures, untested frameworks, untested model families, or tree-speculation variants such as EAGLE and Medusa.

General-purpose large language models (LLMs) are increasingly used for mental health-related conversations, yet safety guardrails remain inadequate and inconsistent across clinical conditions. This study evaluates eight proprietary LLMs across 16 DSM-5 conditions using four adversarial attack variants, introducing an eight-dimension harm taxonomy and a multi-dimensional evaluation framework. Results show that safeguards hold reliably only for suicide and self-harm, while conditions such as eating disorders, substance use disorder, and major depressive disorder exhibit failure rates of up to 100%. We argue that ethical design and deployment of these LLMs demand clearly defined harm categories across clinical conditions and implementation of safeguards accordingly. Until such safeguards are in place, these models pose significant risks to vulnerable populations, making their growing integration into publicly available settings (e.g., schools, search engines, and consumer chatbots) are particularly concerning.

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

Vision-language models (VLMs) are increasingly deployed in consumer, medical, financial, and enterprise applications. This broad deployment expands the safety surface: risks can arise from multimodal question answering, assistant responses, and cross-modal composition, while moderation policies may vary across products, regions, and deployment stages. Most existing guardrails either rely on fixed taxonomies or target only a narrow set of interaction settings, which limits their adaptability when safety rules change at deployment time. We present \textbf{SingGuard}, a policy-adaptive multimodal guardrail model family for safety assessment in multimodal conversations. SingGuard treats the active policy as a runtime input: given natural-language rules, it checks the target content against the active policy rule by rule and predicts both the safety label and the triggered rule. To balance efficiency and interpretability, SingGuard supports fast, hybrid, and slow inference regimes along a fast-to-slow reasoning spectrum, ranging from direct safety judgments to policy-grounded deliberation. We further optimize this behavior with fast--slow decoupled reinforcement learning. We also introduce \textbf{SingGuard-Bench}, a multimodal guardrail benchmark with 56{,}340 examples spanning 80+ fine-grained risk types across multimodal QA, adversarial attack, and dynamic-rule evaluation settings, including cross-modal joint-risk cases where each modality is harmless in isolation but their composition implies unsafe intent. Across six benchmark families (35 datasets), SingGuard achieves state-of-the-art average F1 in every family. Dynamic-rule evaluation further shows improved policy-following accuracy from 0.6465 to 0.7415 under runtime policy shifts. Our code is available at https://github.com/inclusionAI/Sing-Guard.

As Large Language Models (LLMs) achieve widespread integration across diverse linguistic landscapes, ensuring their safety and alignment with regional normative values remains a critical challenge. Current safety mechanisms are predominantly optimized for English-centric frameworks, often failing to capture the unique socio-cultural sensitivities and localized categories of harm inherent to the Indic region. To address this gap, we introduce IndicGuard, a multilingual safety guard model and dataset for Indic languages. We construct a high-volume, culturally nuanced safety dataset encompassing ten major Indic languages, systematically curated to capture regional harms, sensitive socio-political contexts, and adversarial jailbreaks. Leveraging this corpus, we fine-tune a 4B-parameter instruction-tuned model based on Gemma-3-4B-IT to serve as a multilingual safety guardrail for real-time content moderation and policy compliance checking. Our empirical evaluations demonstrate that IndicGuard significantly enhances LLM robustness against localized vulnerabilities, achieving high moderation consistency across different conversational turns. Crucially, IndicGuard consistently outperforms the existing baseline model, CultureGuard, across evaluated languages. Finally, we demonstrate that our model effectively generalizes to low-resource Indic languages excluded from training, substantiating the structural robustness and cross-lingual transfer capabilities of the framework.

Modern Large Language Models (LLMs) rely on extensive safety alignment, yet the mechanistic basis of refusal remains opaque. In this work, we investigate whether safety compliance is a deep semantic decision or a manipulable linear feature. We introduce Contrastive Logit Steering (CLS), a zero-optimization framework that isolates the "refusal direction" by contrasting hidden states derived from safe and unrestricted system prompts. Unlike representation engineering methods that intervene on internal activations, CLS operates directly on the output distribution, serving as a diagnostic probe for alignment fragility. When coupled with prefix injection to bypass initial refusal reflexes, this method induces a phase transition where guardrails collapse. Our experiments on 7 model families reveal that safety implementation is architecturally deterministic. While models like Llama-3.1 exhibit a "Late Decision" topology that is easily bypassed by CLS (reaching 95% ASR in approximately one second), others like Qwen-2.5 demonstrate "Early Divergence" by integrating safety mid-computation. Direct comparison with established activation-level steering methods shows that CLS achieves substantially higher attack success rates on Llama 2 (73% vs. 22.6%) and Qwen 7B (91% vs. 79.2%), demonstrating that logit-level intervention exposes alignment vulnerabilities that hidden-state methods underestimate. Beyond attacks, we show that this linearity enables bidirectional control: inverting the steering vector "hardens" models against jailbreaks without retraining. Our findings suggest that current alignment techniques create a steerable "safety axis" that serves as both a critical vulnerability and a precise primitive for defense.

Coding agents based on large language models (LLMs) demonstrate remarkable autonomous capabilities, but they also introduce significant safety and misuse risks during multi-turn interactions with external environments. Existing safety mechanisms mainly rely on external guardrails, which have a limited ability to perform fine-grained behavioral control during execution. Meanwhile, recent mechanistic interpretability methods for LLM safety are mostly confined to single-turn or jailbreak-style QA settings, limiting their ability to capture the evolving risk dynamics of multi-turn agent execution. In this paper, we investigate the safety of multi-turn coding agents from an internal perspective. We propose AgentLens (Mechanistic Subspace Intervention and Steering), a white-box defense framework that performs runtime safety detection and representation-level mitigation for coding agents. Unlike conventional agent guardrails, AgentLens detect harmful execution states from step-level hidden representations and mitigate unsafe behavior by intervening in a 10-dimensional subspace within a single layer. To support this research, we introduce the Mechanistic Agent Safety (MAS) benchmark, comprising comprehensively annotated multi-turn execution trajectories across 194 tasks using LLaMA-3.1-8B, Qwen-2.5-7B, and Gemma-2-9B. Extensive experiments show that AgentLens achieves strong safety detection performance, provides preliminary evidence for lookahead risk anticipation, and substantially reduces harmful actions of the coding agent, establishing a foundation for applying mechanistic interpretability to dynamic LLM agent safety. The code is available at: https://github.com/EddyLuo1232/AgentLens

Modern LLM agents increasingly rely on context compaction, summarization, or eviction to keep long-running sessions within a token budget. We show that this context-management layer is a safety-critical failure surface: in-context governance constraints that agents reliably obey while visible can be silently removed by compaction, causing the same agent to perform prohibited tool actions later in the session. We call this failure mode Governance Decay. We introduce ConstraintRot, a benchmark of long-horizon agent scenarios with deterministic tool-call grading, and measure compaction-induced violations across seven model families. Across 1,323 episodes, violation rises from 0% with the policy in full context to 30% after compaction, reaching 59% for some models; when the constraint survives the summary, violation remains 0%, but when it is dropped, violation reaches 38%. We further study a Compaction-Eviction Attack, in which adversarial in-context content biases the summarizer to omit a legitimate policy, and show that optimized injections defeat every evaluated model. Finally, we propose Constraint Pinning, a simple training-free mitigation that quarantines governance constraints from lossy compaction and restores violation to 0% in our benchmark. These results identify context management as a first-class governance surface for deployed LLM agents.

Memory safety vulnerabilities remain a significant threat even for projects with extensive fuzzing and manual auditing. Recent results suggest that large language models hold great promise for detecting such vulnerabilities, but they are unreliable, at risk of hallucination, and challenging to scale to repository-size codebases. This paper presents Revelio, a cost-efficient end-to-end agentic framework for memory-safety vulnerability discovery. Revelio addresses the problem of hallucination by generating an executable Proof-of-Vulnerability, which is checked with a deterministic sanitizer. It reduces cost using inexpensive LLMs and lightweight static analysis to help generate and rank vulnerability hypotheses, reporting vulnerabilities only when they can be reproduced and confirmed by a sanitizer. We evaluated Revelio on seven production-quality projects that had been continuously fuzzed for five to eight years, as well as on 100 randomly selected Arvo projects from the CyberGym benchmark. With around one hour per project and a total cost of $300, Revelio discovered 19 previously unknown memory-safety vulnerabilities. On benchmarks, Revelio outperformed frontier coding agents across diverse backbone models at comparable token costs. Our results suggest that Revelio enables scalable and trustworthy end-to-end LLM-based memory-safety vulnerability detection.

Existing safety mechanisms for multimodal large language models (MLLMs) face a fundamental trade-off between safety and utility. Model fine-tuning achieves robust safety but compromises general utility. Input-side safety guardrails offer a lightweight alternative, yet they suffer from severe over-refusal, indiscriminately blocking benign queries or those the model could have safely answered through refusal or advisory responses. We identify that the root cause of over-refusal lies in the input-aware paradigm: safety guardrails make safety decisions without considering whether the model itself is capable of generating safe responses. Usually, MLLMs already possess intrinsic safety mechanisms that can transform harmful inputs into harmless outputs, but input-side safety guardrails override this capability, degrading user experience. Motivated by this insight, we propose a paradigm shift toward output-aware safety guardrails. Our method operates within the model's hidden state space to predict whether the forthcoming generation will be unsafe before it is fully produced. By training a lightweight classifier via multi-instance contrastive learning on hidden state representations, our approach distinguishes between inputs that will lead to unsafe outputs and those that will not, even when the inputs themselves contain risky elements. This enables precise intervention only when the model's actual response would be harmful. Extensive experiments demonstrate that our output-aware safety guardrail matches the safety performance of existing methods while drastically reducing over-refusal, preserving the model's utility and built-in safety capabilities. Code is available at: https://github.com/kunzhan/OutGuard

The increasing deployment of large language model (LLM) agents in collaborative workflows demands robust multi-user, multi-principal interaction mechanisms capable of enforcing access permissions, resolving authoritative conflicts, and preventing unauthorized data disclosure. However, a fundamental mismatch exists between the single-user training paradigm of contemporary LLMs and the hard constraints required for multi-principal governance, rendering probabilistic, prompt-based safeguards vulnerable under multi-turn adversarial interactions.Our key insight is that governance constraints -- who is authorized, what is restricted, and whose instructions take precedence -- are deterministic runtime variables that should be enforced by execution hooks rather than entrusted to the LLM. We present \textbf{Harness-MU}, the first model-agnostic, zero-tuning infrastructure framework for multi-user LLM agents. By decoupling language generation from safety orchestration, Harness-MU guarantees unbreakable permission boundaries while maximizing compliant demand satisfaction. Across four frontier open-weight and proprietary models on the \textit{Muses-Bench} benchmark, Harness-MU achieves the goal of privacy preservation across all access-control attacks, outperforming the standard baseline by 0.28--0.39 in utility score and improving instruction-following accuracy by up to 48.9 percentage points. Harness-MU advances the philosophy of \textit{Harness Engineering}, establishing that systematic infrastructure is essential for solving LLM multi-principal governance challenges. The code and data are available at https://github.com/YuanJrShiuan/Harness-MulUser.

Memory safety errors account for a large proportion of security bugs in systems written in C; modern languages such as Java and Rust prevent such bugs because they are memory-safe by design. To migrate systems to safer languages or identify memory errors, we must first determine how legacy code manipulates memory. This information is only represented implicitly in such code. In many cases, memory usage patterns are merely tedious for humans to figure out, rather than truly difficult. In this work, we ask if large language models (LLMs) can perform this task by having them synthesize annotations representing memory usage as specifications in CN, a hybrid testing/verification tool. Our tool, CNnotator, uses LLMs to automatically generate and test CN specifications. We find that current models are able to generate CN specifications for small-to-medium C programs, with the OpenAI o3 reasoning model achieving a 90% success rate on first attempts and 97% overall success, while the chat model GPT-4o correctly annotates 65% of first attempts. These results suggest AI-assisted annotation is becoming practical for real-world C codebases.

Large Audio Language Models (LALMs) have demonstrated strong performance across a wide range of audio tasks. As they are increasingly deployed in real-world applications, ensuring their safety alignment has become more important. Although refusal mechanisms serve as a key safeguard by preventing LALMs from responding to harmful requests, they can also lead to {\em over-refusal}, where models incorrectly reject benign queries. This issue is especially challenging in the audio domain because speech that appears harmful in isolation may become benign when interpreted together with the surrounding acoustic context, such as background sounds. To study this problem, we introduce \textbf{AOR-Bench} (\textbf{A}udio \textbf{O}ver-\textbf{R}efusal \textbf{Bench}mark), the first benchmark for over-refusal specifically designed for LALMs. AOR-Bench contains 3,000 pseudo-harmful audio samples across six scenario categories. Evaluating 12 representative LALMs from six major model families, we find that over-refusal is widespread (Figure~\ref{fig:overall_performance}) and uncover several important patterns in their safety judgments. As a preliminary effort to mitigate this issue, we further explore two lightweight strategies (e.g., Chain-of-Thought and activation steering) to reduce over-refusal.

Prior work has shown that in-context demonstrations can jailbreak language models, but it remains unclear how models interpret different types of compliance demonstrations. We study this by mixing benign compliance demonstrations (non-harmful request, helpful response) with harmful compliance demonstrations (harmful request, helpful response) and testing three hypotheses about how demonstration composition drives harmful compliance. Across four models, we find that benign and harmful demonstrations are not interchangeable: benign demonstrations can either reduce or increase harmful compliance depending on the model. We further show that preference optimization is the critical training stage that prevents benign demonstrations from increasing harmful compliance, that demonstration ordering exhibits strong recency bias, and that models differ in how refusal interacts with in-context learning: some adopt demonstrated formatting even when refusing, while others override all in-context signals upon refusal. Taken together, this work moves beyond showing that demonstration-based jailbreaking works to characterizing how it works: what models extract from compliance demonstrations depends on demonstration content, ordering, and training methodology.

Whether LLMs scoring well on vulnerability benchmarks genuinely reason about security or merely pattern-match on contaminated data remains unresolved. We present CWE-Trace, a framework for LLM vulnerability detection built from 834 manually curated Linux kernel samples spanning 74 CWEs. The framework enforces a strict temporal split (pre-2025 historical set / post-cutoff leakage-free set), preserves context-aware vulnerable--patched pairs, and introduces two diagnostic metrics: the Directional Failure Index (DFI) and Hierarchical Distance and Direction (HDD). We evaluate eight vanilla LLMs and 15 LoRA fine-tuned variants across non-targeted detection, targeted detection, and CWE classification. Our analysis yields two key results. First, data contamination provides no measurable advantage. Function-level analysis shows that 84% of nominally contaminated samples carry no usable memorization signal: vulnerable functions are absent or cross-mapped across datasets, and ~31% of contaminated samples carry CWE misclassification. Second, backbone directional priors dominate fine-tuning. Models exhibit stable, systematic failure modes (DFI ranging from -85.5 to +94.8 pp) that persist from historical to post-cutoff data and resist correction. Fine-tuning shifts the output threshold without changing the decision policy. This is calibration without comprehension: output distributions adapt to training data while the underlying security reasoning remains absent. The weakest backbone at binary detection (DeepSeek-R1) gains the most in coarse CWE classification, revealing that detection and understanding are decoupled capabilities. The best detection score reaches only 52.1% (+2.1 pp above chance); exact CWE ranking remains below 1.3% Top-1 accuracy, confirming that current LLMs lack reliable security reasoning for systems software, regardless of fine-tuning strategy.

Fine-tuning language models on insecure code induces emergent misalignment with poorly understood internal structure. We investigate whether this misalignment corresponds to a causally actionable activation-space direction shared across architectures. Across four instruction-tuned model families (Qwen2.5-1.5B, Gemma-2-2B, Llama-3.2-1B, Ministral-3-3B) finetuned identically, a difference-in-means direction achieves 99.6% separation of aligned and misaligned activations at each model's final layer. Causal steering by subtracting this direction reduces code spillover by 21-51 points, while a secure-code control confirms content specificity. Cross-architecture transfer via ridge regression maps yields large behavioral suppression (up to 46 points) but fails specificity controls as random and orthogonal directions perform comparably. We identify a two-tier specificity structure: within-model directions are causally specific and actionable; cross-model directions are causally real but non-specific. An asymmetric transfer topology emerges, with Gemma and Qwen acting as geometric donors and Llama as a receiver. These findings define the limits of linear cross-architecture correction and recommend within-model probing for auditing.

Existing safety benchmarks target general adversarial scenarios but miss finance-specific risks. Financial LLMs face regulatory compliance violations, fraud facilitation, and systemic trust erosion that require targeted evaluation. We introduce FinRED, an expert-guided red-teaming framework for financial LLM safety evaluation developed with financial experts. FinRED uses a novel two-level taxonomy mapping global standards (e.g., FATF and EU DORA) to threats ranging from regulatory evasion to complex fraud, integrated with a scalable pipeline that converts real financial documents into context-rich red-teaming Behavioral Prompts (seeds) through an expert-defined schema. Rigorous expert validation confirms seed plausibility and realism for meaningful LLM safety evaluation. We also provide an expert-validated, finance-specific rubric that goes beyond disclaimer checks, aligns more closely with human experts than static one-size-fits-all rubrics, and reduces critical false negatives from 28 to 12. Aligned with internationally adopted risk-management and information-security standards (e.g., ISO/IEC 27001), FinRED is deployed in South Korea's Financial Security Institute (FSI) regulatory sandbox for generative AI security evaluation in real financial services. To mitigate dual-use risks, the dataset, generation pipeline, prompt template, and evaluation framework are gated for qualified researchers at https://github.com/selectstar-ai/FinRED-paper and https://huggingface.co/datasets/datumo/FinRED.

Speculative inference accelerates large language model (LLM) decoding but provides no inherent safety guarantees. Existing safety defenses are largely incompatible with speculative inference: they either introduce additional computation or disrupt the draft-verify mechanism, negating acceleration benefits. This reveals a fundamental incompatibility between current safety methods and speculative decoding. We propose SafeSpec, a safety-aware speculative inference framework that integrates risk estimation directly into the verification process. SafeSpec attaches a lightweight latent safety head to the target model to jointly evaluate semantic validity and safety in a single forward pass. When unsafe generations are detected, SafeSpec applies rollback and safety-guided reflective multi-sampling to recover safe continuations rather than terminating generation. We model jailbreak attacks as distributional shifts over generative trajectories, where adversarial prompts increase the probability of harmful continuations without eliminating safe ones. Under this model, SafeSpec performs risk-aware trajectory recovery within the speculative decoding process. Across multiple models and adversarial benchmarks, SafeSpec achieves a substantially improved safety-efficiency trade-off. On Qwen3-32B, SafeSpec reduces attack success rates by 15% while preserving a 2.06x inference speedup on benign workloads, demonstrating that speculative acceleration and inference-time safety can be jointly optimized.

Formal tools such as SAT and SMT solvers are increasingly embedded in language model reasoning pipelines when a safety or security critical question can be formulated in logic. Unlike chain of thought whose steps are sampled from the model distribution without formal guarantee, a solver produces a sound and independently verifiable answer. However, the soundness guarantee can be lost in the interaction between the solver and the model. The hybrid pipeline has three components: formalizing the question, deciding it, and narrating the result. Prior work has studied the formalization and decision, but not narration, which is the step that turns a formal tool's output into the user answer. To fill the narration gap, we first model the LLM-solver loop as a verified decision procedure. We further evaluate five open-sourced models under prompt injection, and we find certificate gating makes the solver verdict sound, while an adversary can invert a verified conclusion across phrasings and channels. We study the mitigation through hardened prompt that reduces injection significantly but cannot eliminate it and still suffers under adaptive attack. Combining the formal analysis and empirical studies, we show in the LLM-solver loop, robustness does not reach to the answer that the user finally reads.

Large language models (LLMs) are increasingly embedded in AI for Science (AI4Science) workflows, from scientific question answering and literature analysis to laboratory planning and autonomous discovery. This progress creates an urgent need for safety benchmarks that evaluate not only scientific competence, but also whether models recognize and avoid risks in high-stakes scientific contexts. Existing AI4Science safety datasets cover several disciplines and task formats, leaving the underlying risk dimensions underspecified. We introduce \textbf{SciRisk-Bench}, a benchmark designed to evaluate AI4Science safety from two complementary perspectives: explicit risk dimensions and scientific disciplines. SciRisk-Bench covers 7 disciplines, 31 subdisciplines and 10 risk dimensions. In the experimental section, we evaluate both mainstream LLMs and science-oriented LLMs across risk dimensions, disciplines, and sub-disciplines, enabling fine-grained diagnosis of where scientific models remain unsafe.

Tool-using language-model agents introduce security failures that go beyond unsafe text: they can disclose protected objects, write persistent memory, send messages, modify databases, or trigger harmful code and tool effects. Existing evaluations often collapse these stages into a single attack success rate, making it difficult to tell whether a model merely agreed with an attacker or actually produced observable harm. We introduce SafeClawBench, a staged benchmark for tool-using agent security with 600 controlled adversarial tasks across six attack families: direct and indirect prompt injection, tool-return injection, memory poisoning, memory extraction, and ambiguity-driven unsafe inference. SafeClawBench reports three separate endpoints: semantic attack acceptance, audit-visible harm evidence, and sandbox-observed tool/state harm. Evaluating five agent endpoints under four prompt-level policies, we find that these endpoints capture different failure modes. Without additional prompt protection, semantic failure rates vary widely across models, from 9.0% to 44.2%. Audited harm evidence is narrower than semantic failure, and under a separate executable protocol some matched task identities produce sandbox harm despite passing the Semantic Core call: in a 12,000-row matched analysis, 291 of 347 observed sandbox harms occur in rows that pass the semantic check. Prompt policies change endpoint outcomes, but their effects depend on both model and protocol. SafeClawBench provides a reproducible framework for comparing agent models and prompt-policy conditions without conflating textual compliance, evidence-supported harm, and executable state changes. The open-source dataset is available at https://huggingface.co/datasets/sairights/safeclawbench.

Large language models (LLMs) outperform earlier architectures on generative inference and long-context tasks, but their large size introduces significant challenges in memory usage, energy cost, and on-device deployment. Since scaling pre-trained language models improves downstream capability \cite{zhao2023survey}, the key-value (KV) cache becomes a dominant inference bottleneck. Recent KV cache compression methods \cite{jo2025fastkv,li2024snapkv,zhou2024dynamickv} reduce this cost by retaining only a subset of attention-relevant tokens. However, while these approaches preserve accuracy on benign workloads, their compression policies either fail to defend against jailbreak attacks \cite{jiang2024robustkv} or degrade safety alignment under aggressive eviction. We propose AnchorKV, a drop-in modification to KV cache compression that biases token retention scores away from directions in key space associated with harmful prompts. AnchorKV constructs an offline safety anchor by adapting a difference-of-means representation engineering approach \cite{arditi2024refusal,zou2023representation} to the layer-specific key projection space used in KV caching. Based on this anchor, a soft penalty token selection rule trades a small amount of utility for substantially improved safety alignment, while reducing to the original compressor when the penalty is zero.

While Large Reasoning Models (LRMs) excel at complex tasks, they remain highly vulnerable to sophisticated jailbreaks and direct harmful queries. To address this vulnerability, prior works depend heavily on external manual data annotation for safety alignment. However, we observe that LRMs can inherently identify safety risks when being re-presented with original queries alongside their own reasoning trajectories -- a capability we term Latent Safety Awareness. To leverage this safety awareness, we first employ Supervised Fine-Tuning (SFT) to explicitly induce safe tags to trigger safety analysis and guidance following the initial reasoning content for unsafe queries, while preserving standard responses for general queries to ensure adaptive triggering. Subsequently, we apply Direct Preference Optimization (DPO) to further enhance the correctness and stability of the safety analysis and guidance. Notably, responses required for both training stages are entirely generated by models being optimized. With (Safe Trigger) SFT and DPO, experimental results demonstrate significant safety enhancement. For example, the Attack Success Rate (ASR) of DeepSeek-R1-Distill-Llama-8B, on average, drops 24.65% and 36.72% on harmful and jailbreak benchmarks, respectively. Finally, our Safe Trigger method exerts almost no negative impact on general performance or user experience.

Recent work motivates moving large language model (LLM) evaluation from mean-based to tail-aware metrics, including conditional value-at-risk and tail-index estimates of reward-model error. We ask whether the canonical extreme-value-theory tail-index parameter, which isolates how heavy a tail is from how large the tail mass is, adds discriminative information beyond the mean and a standard tail-magnitude statistic in LLM evaluation. We pre-register a protocol covering admissibility, goodness-of-fit, threshold-stability, and effect-size requirements for any positive tail-shape claim. The protocol is the contribution of this paper; the empirical study below is a demonstration of what its gates catch. Applied to a standard LLM toxicity-evaluation setup under two structurally different scorer families, the protocol catches three distinct modes of false positives that a naive analysis would have published, and rejects the headline tail-shape claim on both scorers. We conclude that tail-shape estimation in the LLM toxicity-evaluation setups we examined is more fragile than the recent literature suggests, and recommend the protocol as a starting point for tail-index claims in similar setups.

Large language models routinely generate code with exploitable security flaws. Prior literature attributes this limitation to a lack of security expertise, steering current defense mechanisms toward heavy fine-tuning or external knowledge retrieval, which introduces significant computational overhead and data bias through redundant code examples. Contrary to this view, we argue that pretraining corpora are already rich in security material. The bottleneck is activation: without an explicit and brief cue, statistical pressure toward common training-distribution patterns suppresses the model's safety-relevant representations. We present SPARK, an inference-time security harness that activates this latent knowledge without any retraining. The harness has two parts. ComponentI retrieves a few of the relevant Common Weakness Enumeration (CWE) entries for each coding task and appends a short structured cue to the prompt; this alone is enough to surface the model's existing security representations. ComponentII adds a precomputed token bias to the logits at every decoding step. We obtain the bias by projecting a safe-direction vector, the unit difference between the mean safe and mean unsafe last-layer hidden states, through the language model head. The bias is computed once offline; applying it costs a single vector addition per generated token. We evaluate SPARK on 9 open-source models across C++, Java, and Python, and compare with 7 baselines spanning fine-tuning and retrieval-augmented methods. SPARK matches or improves on the best baseline in every setting while preserving HumanEval utility. We further test Component~I in a black-box setting on 7 of today's strongest models, including Claude, DeepSeek, and GPT, demonstrating the bottleneck of insecure code generation and the improvements enabled by our method.

Open-source LLM agent ecosystems are growing rapidly, yet the security of community-contributed skills - modular tool definitions that extend agent capabilities - remains largely unvetted. The gap we fill: existing scanners operate at the code layer and are structurally blind to instruction-layer and multi-agent risk - natural-language directives that hijack an agent, exfiltrate data through encoded side channels, or chain harm across pipelines - so what is needed is a semantic, multi-dimensional vetting system rather than another signature matcher. We present SKILLVETBENCH, a live public leaderboard on Hugging Face that uses an LLM-as-Judge to vet agent skills. What is new: SARS (Skill Agentic Risk Score), a five-dimensional agentic-risk metric with a principled weighted formula for instruction-following systems. What is integrated: full CVSS v4.0 vector decomposition and a ClawHub dual-view that places our LLM-generated review beside the official marketplace verdict. What is demonstrated: drawing on our companion benchmark paper [ 1], the LLM-as-Judge stage achieves zero false negatives across 78 confirmed-malicious skills and zero false positives across 22 benign controls, while the best static baseline (SKILLSIEVE) still misses 15%; for instruction-layer categories such as Prompt Injection and Memory Poisoning, conventional tools miss between 89% and 100% of threats (e.g., CODEBERT detects none of nine memory-poisoning skills). Detection rates vary from 35% to 95% across four LLM evaluators, motivating ensemble scoring in production deployments.

We ran 300 repeated vulnerability-finding scans to measure how repeatable agentic large language model (LLM) security review is on the same JavaScript code, prompt, and benchmark harness. The headline result is that LLM security findings were unevenly repeatable: reference-matched findings were stable, but extra model reports varied heavily from run to run. Across 250 model runs, 80 of 161 unique unmatched findings appeared in only one of five identical repetitions, while only 22 appeared in all five. By contrast, when Claude matched a Snyk Code reference finding, the behavior was much more stable: 134 of 158 unique reference-matched findings appeared in all five repetitions. The benchmark also shows complementarity. Models consistently found familiar, high-signal exploit shapes, and in one case surfaced a likely Snyk Code product gap. Snyk Code static application security testing (SAST) was deterministic and better at systematically enumerating repeated data-flow sinks. The results support combining agentic LLM review with deterministic SAST rather than treating either technique as a replacement for the other.

Large language models often struggle with sensitive prompts. They may refuse outright, provide generic safety boilerplate, or fail to address the user's legitimate informational needs that can be answered safely. We introduce SHARD, a self-reframing distillation method to improve safe-helpfulness. It first rewrites sensitive prompts to surface benign intent using philosophical guidelines, then reframes its original responses into safe, more helpful ones, and finally fine-tunes the model on its self-reframed responses. Across DNA and the English subset of LINGUASAFE, SHARD improves helpfulness for most model families while preserving safety. It also remains competitive with distillation from a larger teacher model, suggesting that models can internalize safe and helpful behavior elicited from their own. Warning: This paper contains content that may be offensive or harmful.

Malicious content generated from large language models (LLMs) could pose severe safety risks and ethical concerns. While existing LLM safety guardrails excel in English or multilingual settings, they lack adaptation to Chinese-specific regulatory policies, cultural context and linguistic nuances, failing to support fine-grained risk classification for diverse deployment needs. In this paper, we introduce a 5-macro, 31-micro category fine-grained risk taxonomy for Chinese scenarios, and build CHILLGuard: a dedicated Chinese LLM content safety guardrail. To address the critical scarcity of high-quality annotated Chinese safety data, we propose a scalable multi-stage data construction pipeline: we expand multi-source corpus via retrieval-augmented generation, generate implicit harmful samples through prompt engineering rewriting, and refine high-quality data via multi-model voting-based label calibration. Based on this, we build CHILLGuardTrain, a large-scale training set with 405,007 samples, and CHILLGuardTest, a rigorously curated annotated test set with 51,745 samples. We then train CHILLGuard on CHILLGuardTrain under a generator-classifier collaborative framework via Model-aware Direct Preference Optimization. Extensive experiments under multiple settings demonstrate the state-of-the-art performance of CHILLGuard, e.g., a 15.92% improvement of F1 score over Qwen3Guard-8B-Strict on our benchmark. We will release our resources at https://github.com/cswbyu/CHILLGuard.

This paper appraises recent frameworks within AI development to integrate LLMs into control tasks in automotive contexts from the perspective of safety assurance. This work has built upon the rapid integration of LLMs across automotive settings. However, we find that at present, these frameworks face significant challenges, limiting their efficacy in real-time safety-critical contexts. Firstly, we consider conceptual challenges, including the fact that deployers are faced with a dual challenge, wherein they must assure a model which has been developed upstream, i.e. as general-purpose tools by the large AI labs, in a downstream context, i.e. into specific vehicle architectures. Secondly, we consider concrete challenges from across existing standards. We show that there are currently both fundamental engineering constraints covered in ISO21448, such as latency, and novel LLM-specific issues, such as alignment-related issues covered in ISO/PAS8800. We ground both examples in a concrete introductory, experimental case study exploring an existing open-source repository, Talk2Drive. We present a safety argument in order to make explicit the limitations of existing solutions. Nonetheless, given that the use of LLMs in automotive contexts is being explored at a technical level and operationalised, we propose potential assurance mechanisms for LLM-related hazardous events going forward.

Large Language Models (LLMs) are increasingly deployed in healthcare settings, yet their tendency to hallucinate poses risks when clinical decisions are involved. This study examine whether LLMs recommend recently banned or withdrawn pharmaceuticals when answering clinical questions and tests an agent-based method for reducing such errors. We developed a five-agent "Trust but Verify" system using a single LLM backbone. To measure regulatory knowledge obsolescence, we created an adversarial dataset of 103 clinical MCQs where historically correct answers now refer to banned substances. This scale ensures statistical significance across various therapeutic classes. We evaluated three open-access model families (GPT-OSS, Llama-3, Falcon-3) under vanilla and agentic conditions. Performance was measured via pointwise score, label accuracy, Hallucination Error Rate (HER), and Component Fidelity (CF) score. We also observed clinical safety regression in proprietary models. In default configurations, all models showed high hallucination rates, consistently selecting banned drugs that matched training data patterns. Our proposed agentic architecture reduced HER by approximately 53% across models. Pointwise scores shifted from -0.25 (unsafe recommendation) toward 0.0 (appropriate refusal). The safety audit intercepted dangerous outputs even when models' parametric knowledge favored the banned substance. The proposed multi-agent framework offers a model-agnostic method for enforcing regulatory compliance that prioritizes patient safety over fluent text generation. Our work demonstrates a practical approach for deploying autonomous AI systems in safety-critical healthcare settings. It shows how real-time regulatory data can be integrated into LLM pipelines to support clinical decision-making.

We present the design and implementation of a safety constrained large language model (LLM) system for public health information access, focusing on maternal and child health (MCH) resource navigation. While LLM based systems offer flexible and natural interfaces for information retrieval, their deployment in healthcare contexts introduces risks related to safety, trust, and uncontrolled generation. This work explores practical design patterns for constraining LLM behavior in safety critical environments. We introduce a multi-layered architecture that integrates domain-restricted retrieval augmented generation (RAG), strict boundary enforcement to prevent medical advice, anonymous multiuser session management, and comprehensive audit logging for monitoring and compliance. A key aspect of the design is a controlled data pipeline that grounds all responses in curated public health resources, avoiding reliance on the model pretrained medical knowledge. We implement the system in a real world public health setting and conduct scenario-based validation across in scope, out of scope, and emergency queries. Results show consistent enforcement of safety constraints, reliable resource grounding, and stable system performance, with an average response time of 5.3 seconds. Beyond the specific application, we discuss design trade offs and lessons learned in balancing safety, usability, and system flexibility. Our findings provide practical guidance for deploying LLM based systems in healthcare and other domains where strict information boundaries and accountability are required.

Stream guardrails enable token-level safety detection before full responses are generated. However, they often make overly conservative judgements and block those sensitive but safe tokens, which is known as over-refusal. Due to lack of full context, they also fail to detect implicitly harmful content from jailbreaking. To address these challenges, we propose FreoStream, a novel streaming guardrail framework. Specifically, FreoStream fine-tunes a LoRA module to perform Future-Aware Reasoning when the base guardrail detects unsafe tokens. The reasoning process follows a Future-Reason-Judge paradigm: predict the future, reason about the full context and give the final judgement. This design can effectively reduce over-refusal by incorporating the future information. Moreover, we introduce the Safety-Aligned Optimization module that extracts the safety-aligned component from the reasoning gradients to update the base guardrail model, thereby enhancing streaming safety detection. Extensive experiments on various safety benchmarks demonstrate that FreoStream achieves lower over-refusal rates and better jailbreak defense compared to existing streaming guardrails.

Large language models (LLMs) are increasingly used to access organisational documentation, including standard operating procedures (SOPs), HR policies and institutional guidelines. However, retrieval-augmented generation (RAG) systems that rely on free-form rewriting can introduce hallucinations and unstable trade-offs between completeness and conciseness, particularly in safety- and compliance-critical settings. Objectives: To evaluate extraction as a hallucination-resistant alternative to rewriting-based RAG and compare strategies that balance precision, recall and safety across document types and model scales. Methods: We compare multiple prompting strategies, including line-number-based source selection, extraction of relevant guideline sentences with explicit safety annotations, and a multi-stage pipeline that refines draft answers using supporting evidence from source guidelines. Experiments are conducted on documents of varying length and structure, including local NHS acute care and oncology guidelines and UK-wide NICE guidelines, using both frontier-scale and locally deployable models. Performance is assessed using automatic metrics and human expert evaluation of relevance and completeness. Results: Line-number selection achieves the strongest results, outperforming direct copying and safety-focused strategies across both large and small models while maintaining high term recall (up to 95%) and close alignment with source text. Safety-oriented approaches improve precision but introduce systematic omissions, while multi-stage filtering further amplifies this trade-off. Performance varies with document structure: line-based extraction excels in protocol-like content, whereas alternative strategies perform better on more verbose documents (up to 97% term recall).

Agentic large language model systems that autonomously invoke tools, maintain persistent memory, and execute multi-step plans are increasingly deployed in public-facing domains, including government services, healthcare triage, and financial advising. We ask whether the frameworks used to build these systems provide architectural-level structural safety guarantees. Applying six containment principles derived from a compositional model of agentic architectures, we audit three dominant frameworks (LangChain, AutoGPT, and OpenAI Agents SDK) and find no native compliance in any of them. Memory integrity, a defense against one of the most prevalent vulnerability classes, is not observed in any of the three evaluated frameworks. We validate these findings empirically: in a simulated government benefits agent built on LangChain, a single memory-poisoning write induces persistent targeted corruption across all tested seeds and backends, increasing the wrongful denial rate for targeted applicants to 88.9%. Under a complex five-factor policy, the same attack preserves aggregate accuracy while increasing targeted wrongful denials by 3.5x, rendering the corruption difficult to detect through standard monitoring. We then introduce two lightweight containment mechanisms: a memory integrity validator and a policy gate, which eliminate both attack vectors with sub-millisecond overhead (<0.2ms per call). We conclude that the current agentic framework ecosystem may not yet meet secure-by-default expectations for public-facing deployments and outline priority architectural interventions to enable trustworthy deployment in high-stakes, socially impactful applications.

Safety-relevant studies of language models, including alignment and jailbreaking evaluations and AI control protocols, often rely on prefilling model outputs. If AI models can recognize and act on the fact their prior assistant messages have been inserted or edited, the effectiveness and validity of these methods could be compromised. We investigate whether frontier language models can distinguish between tampered and untampered assistant-side context, a capability we call prefill awareness. To do so, we construct a binary preference benchmark across three prefill mechanisms, filtering for cases where models show consistent stances. We find that frontier models show substantial prefill awareness: Claude Opus 4.5 detects prefills opposing its preferences in 9-35% of cases with a 0% false positive rate when prompted; additionally, models often revert towards baseline behavior without explicitly reporting that the prefill was foreign. Controlled ablations later also show that detection and resistance rely on different cues, where stylistic mismatch mainly affects whether models flag a prefill as foreign, while preference mismatch mainly affects whether they revert toward their baseline answer. We also examine more realistic agentic settings such as misalignment-continuation evaluations and SWE-bench trajectories, where frontier models sometimes disavow prefilled assistant turns in ways that depend strongly on dataset, task success, and hidden formatting artifacts. Our results indicate that prefill awareness is already a substantial confound for some prefill-based methods. We recommend that model developers track this capability in frontier systems.

Repository-level benchmarks for evaluating Large Language Model (LLM) code repair on Secure Multi-Party Computation (MPC) software do not yet exist, and directly transplanting general-purpose benchmarks such as SWE-bench fails on three structural fronts: (i) MPC repositories are dominated by generic Python infrastructure rather than cryptographic logic; (ii) high-value MPC fixes lack the standardized tests rigid extraction pipelines require; and (iii) standard fail-to-pass evaluation is insufficient for code that must also be cryptographically safe. MPC is increasingly deployed for privacy-preserving machine learning, biomedical collaboration, and secure analytics. Existing MPC-specific code-synthesis efforts cover only operator-level or single-framework tasks; evaluating LLM agents on real repository-level MPC repair instead demands MPC-aware data curation and a verifier matched to the security and numerical-fidelity guarantees MPC programs must obey neither of which existing benchmarks provide. We introduce MPC-Patch-Bench, a repository-level benchmark organised around two frameworks. (1)The Data Curation Framework combines a domain-specific curation agent that filters raw pull requests through three cryptographic layers with a human-AI completion engine that synthesizes missing problem statements and Fail-to-Pass/Pass-to-Pass tests, yielding 205 fully verified instances. (2)The MPC Verifier provides dedicated security and numerical-fidelity checks via dynamic differential testing against plaintext oracles and MPC-specific static analysis rules that flag unsafe reveals, insecure arithmetic, and illegal public/private casts. The strongest evaluated LLM functionally resolves only 22.9% of MPC-Patch-Bench tasks; the MPC Verifier further reduces verified resolution to 17.1%, with up to 40% of functionally-passing patches rejected for cryptographic or numerical-fidelity violations.

Large language models are increasingly deployed across professional domains, bringing hard-to-predict risks, including the generation of harmful or disrespectful content. Although substantial progress has been made in developing safety evaluation datasets, existing resources remain overwhelmingly English- and Chinese-centric. This limitation is particularly pronounced when evaluating languages that operate within shared sociocultural, legal, and ethical contexts. To address this gap, we introduce Schützen: a German--Bulgarian safety dataset designed to assess model answerability under risk, covering both a low-resource language (Bulgarian) and a high-resource language (German). Experiments with multilingual and language-specific LLMs reveal pronounced cross-language differences in safety behavior, highlighting the necessity of tailored, region-specific evaluation resources to support the responsible deployment of LLMs in Germany and Bulgaria. Datasets and code are available at https://github.com/xnlp-lab/Schutzen. Warning: this paper contains examples that may be offensive, harmful, or biased.

Deploying large language models in user-facing systems requires efficient output safety filtering. Existing approaches typically rely on a separate moderation model applied after generation, which doubles inference cost and only detects violations after generation completes. We observe that the signal needed for moderation is already present in the model hidden states. Based on this, we train lightweight token-level probes that operate directly on internal activations, producing per-token safety scores that can be aggregated for both offline evaluation and online intervention. The probe reuses activations from the generator and requires no additional forward pass, enabling sub millisecond per-token safety checks inside the decoding loop. A probe applied to a single mid layer recovers most decisions of a strong guard model, acting as a low cost surrogate optimized for latency rather than accuracy. In streaming settings, it can halt or modify unsafe outputs before they are fully generated, replacing end of sequence moderation with continuous token level monitoring. Compared to post hoc and streaming guard models, our method achieves orders of magnitude lower compute overhead with minimal latency cost. We also provide a practical deployment recipe, including layer selection, aggregation strategy, probing frequency, and triggering thresholds. Finally, we show that the probe linear component corresponds to a direction in residual space, enabling both detection and activation steering at negligible cost.