Large Language Model Safety

Cost-aware routing dynamically dispatches user queries to models of varying capability to balance performance and inference cost. However, the routing strategy introduces a new security concern that adversaries may manipulate the router to consistently select expensive high-capability models. Existing routing attacks depend on either white-box access or heuristic prompts, rendering them ineffective in real-world black-box scenarios. In this work, we propose R$^2$A, which aims to mislead black-box LLM routers to expensive models via adversarial suffix optimization. Specifically, R$^2$A deploys a hybrid ensemble surrogate router to mimic the black-box router. A suffix optimization algorithm is further adapted for the ensemble-based surrogate. Extensive experiments on multiple open-source and commercial routing systems demonstrate that {R$^2$A} significantly increases the routing rate to expensive models on queries of different distributions. Code and examples: https://github.com/thcxiker/R2A-Attack.

As agent systems move into increasingly diverse execution settings, trajectory-level safety evaluation and diagnosis require benchmarks that evolve with them. ATBench is a diverse and realistic agent trajectory benchmark for safety evaluation and diagnosis. This report presents ATBench-Claw and ATBench-Codex, two domain-customized extensions that carry ATBench into the OpenClaw and OpenAI Codex / Codex-runtime settings. The key adaptation mechanism is to analyze each new setting, customize the three-dimensional Safety Taxonomy over risk source, failure mode, and real-world harm, and then use that customized taxonomy to define the benchmark specification consumed by the shared ATBench construction pipeline. This extensibility matters because agent frameworks remain relatively stable at the architectural level even as their concrete execution settings, tool ecosystems, and product capabilities evolve quickly. Concretely, ATBench-Claw targets OpenClaw-sensitive execution chains over tools, skills, sessions, and external actions, while ATBench-Codex targets trajectories in the OpenAI Codex / Codex-runtime setting over repositories, shells, patches, dependencies, approvals, and runtime policy boundaries. Our emphasis therefore falls on taxonomy customization, domain-specific risk coverage, and benchmark design under a shared ATBench generation framework.

Large language models are increasingly used as natural-language interfaces to enterprise software, but their direct use as system operators remains unsafe. Model errors can propagate into unauthorized actions, malformed requests, cross-workspace execution, and other costly failures. We argue this is primarily an execution architecture problem. We present a bounded-autonomy architecture in which language models may interpret intent and propose actions, but all executable behavior is constrained by typed action contracts, permission-aware capability exposure, scoped context, validation before side effects, consumer-side execution boundaries, and optional human approval. The enterprise application remains the source of truth for business logic and authorization, while the orchestration engine operates over an explicit published actions manifest. We evaluate the architecture in a deployed multi-tenant enterprise application across three conditions: manual operation, unconstrained AI with safety layers disabled, and full bounded autonomy. Across 25 scenario trials spanning seven failure families, the bounded-autonomy system completed 23 of 25 tasks with zero unsafe executions, while the unconstrained configuration completed only 17 of 25. Two wrong-entity mutations escaped all consumer-contributed layers; only disambiguation and confirmation mechanisms intercept this class. Both AI conditions delivered 13-18x speedup over manual operation. Critically, removing safety layers made the system less useful: structured validation feedback guided the model to correct outcomes in fewer turns, while the unconstrained system hallucinated success. Several safety properties are structurally enforced by code and intercepted all targeted violations regardless of model output. The result is a practical, deployed architecture for making imperfect language models operationally useful in enterprise systems.

Large language models (LLMs) frequently generate toxic content, posing significant risks for safe deployment. Current mitigation strategies often degrade generation quality or require costly human annotation. We propose CAUSALDETOX, a framework that identifies and intervenes on the specific attention heads causally responsible for toxic generation. Using the Probability of Necessity and Sufficiency (PNS), we isolate a minimal set of heads that are necessary and sufficient for toxicity. We utilize these components via two complementary strategies: (1) Local Inference-Time Intervention, which constructs dynamic, input-specific steering vectors for context-aware detoxification, and (2) PNS-Guided Fine-Tuning, which permanently unlearns toxic representations. We also introduce PARATOX, a novel benchmark of aligned toxic/non-toxic sentence pairs enabling controlled counterfactual evaluation. Experiments on ToxiGen, ImplicitHate, and ParaDetox show that CAUSALDETOX achieves up to 5.34% greater toxicity reduction compared to baselines while preserving linguistic fluency, and offers a 7x speedup in head selection.

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

Vision-Language-Action (VLA) models inherit rich world knowledge from vision-language backbones and acquire executable skills from action demonstrations. Yet current evaluations primarily measure task completion, leaving the semantic safety of learned action policies underexplored. This gap creates a critical vulnerability: a policy may execute the intended action correctly while producing unsafe outcomes when the surrounding visual-linguistic context changes. We present HazardArena, a benchmark for stress-testing semantic safety in VLA systems. Its core design is a set of safe/unsafe twin scenarios: paired environments with matched objects, layouts, and action requirements, but different semantic risk contexts. This controlled contrast isolates safety judgment from motor capability and directly tests whether a VLA can recognize when an otherwise valid action becomes hazardous. HazardArena includes over 2,000 assets and 51 risk-sensitive tasks across seven safety categories grounded in robotic safety standards. Across four representative VLA backbones, we observe a consistent and alarming pattern: safe-only fine-tuning improves benign task success while also increasing hazardous execution on matched unsafe scenarios. Physical-world experiments confirm that this failure transfers beyond simulation. These results show that stronger action execution does not imply safer behavior, and motivate semantic-risk-aware evaluation and enforcement as first-class requirements for real-world VLA deployment. Code released at https://github.com/HazardArena-Team/HazardArena ; updated code availability information.

An OS kernel that runs LLM inference internally can read the model's own next-token logit distribution before any text is generated, and act on it as a governance primitive. I present ProbeLogits, a kernel-level operation that performs a single forward pass and reads specific token logits to classify an agent's action as safe or dangerous, with zero learned parameters. Because the probe reads a logit from the same base model the agent already runs, it removes the second model a fine-tuned guard requires: the marginal cost of a safety check becomes a single logit read. I evaluate ProbeLogits on three base models (Qwen2.5-7B, Llama-3-8B, Mistral-7B) across three external benchmarks (HarmBench, XSTest, ToxicChat). On HarmBench non-copyright, all three reach a 97-99% block rate. On ToxicChat (n=1,000), ProbeLogits attains F1 parity-or-better against Llama Guard 3: Qwen2.5-7B Safe/Dangerous reaches F1=0.812 (+13.7 pp, bootstrap 95% CIs disjoint), Llama-3 matches within CI (+0.4 pp), and Mistral exceeds by +4.4 pp. Classification is a measured 2.4-3.4x faster than Llama Guard 3 (332-556 ms vs. 851-1,142 ms), because it reads a single logit position instead of generating tokens. A calibration strength alpha acts as a deployment-time policy knob rather than a learned hyperparameter, trading recall for precision per operation class. I implement ProbeLogits within Anima OS, a bare-metal x86-64 kernel written in ~285,000 lines of Rust. Because agent actions must pass through kernel-mediated host functions, enforcement operates below the WASM sandbox boundary, making it substantially harder to circumvent than application-layer classifiers.

Large language models (LLMs) undergo alignment training to avoid harmful behaviors, yet the resulting safeguards remain brittle: jailbreaks routinely bypass them, and fine-tuning on narrow domains can induce ``emergent misalignment'' that generalizes broadly. Whether this brittleness reflects a fundamental lack of coherent internal organization for harmfulness remains unclear. Here we use targeted weight pruning as a causal intervention to probe the internal organization of harmfulness in LLMs. We find that harmful content generation depends on a compact set of weights that are general across harm types and distinct from benign capabilities. Aligned models exhibit a greater compression of harm generation weights than unaligned counterparts, indicating that alignment reshapes harmful representations internally--despite the brittleness of safety guardrails at the surface level. This compression explains emergent misalignment: if weights of harmful capabilities are compressed, fine-tuning that engages these weights in one domain can trigger broad misalignment. Consistent with this, pruning harm generation weights in a narrow domain substantially reduces emergent misalignment. Notably, LLMs harmful generation capability is dissociated from how they recognize and explain such content. Together, these results reveal a coherent internal structure for harmfulness in LLMs that may serve as a foundation for more principled approaches to safety.

We localize the policy routing mechanism in alignment-trained language models. An intermediate-layer attention gate reads detected content and triggers deeper amplifier heads that boost the signal toward refusal. In smaller models the gate and amplifier are single heads; at larger scale they become bands of heads across adjacent layers. The gate contributes under 1% of output DLA, yet interchange testing (p < 0.001) and knockout cascade confirm it is causally necessary. Interchange screening at n >= 120 detects the same motif in twelve models from six labs (2B to 72B), though specific heads differ by lab. Per-head ablation weakens up to 58x at 72B and misses gates that interchange identifies; at scale, interchange is the only reliable audit. Modulating the detection-layer signal continuously controls policy from hard refusal through evasion to factual answering. On safety prompts the same intervention turns refusal into harmful guidance, showing that the safety-trained capability is gated by routing, not removed. Thresholds vary by topic and by input language, and the circuit relocates across generations within a family even while behavioral benchmarks register no change. Routing is early-commitment: the gate fires at its own layer before deeper layers finish processing the input. An in-context substitution cipher collapses gate interchange necessity by 70 to 99% across three models, and the model switches to puzzle-solving rather than refusal. Injecting the plaintext gate activation into the cipher forward pass restores 48% of refusals in Phi-4-mini, localizing the bypass to the routing interface. A second method, cipher contrast analysis, uses plain/cipher DLA differences to map the full cipher-sensitive routing circuit in O(3n) forward passes. Any encoding that defeats detection-layer pattern matching bypasses the policy regardless of whether deeper layers reconstruct the content.

Recent advances in vision-language models (VLMs) have accelerated their application to indoor safety hazards assessment. However, existing benchmarks suffer from three fundamental limitations: (1) heavy reliance on synthetic datasets constructed via simulation software, creating a significant domain gap with real-world environments; (2) oversimplified safety tasks with artificial constraints on hazard and scene types, thereby limiting model generalization; and (3) absence of rigorous evaluation protocols to thoroughly assess model capabilities in complex home safety scenarios. To address these challenges, we introduce TSHA (\textbf{T}rustworthy \textbf{S}afety \textbf{H}azards \textbf{A}ssessment), a comprehensive benchmark comprising 66,668 validated question-answer pairs, including 64,961 carefully curated training QA pairs drawn from existing indoor datasets, internet frames/images, AIGC images, newly captured images, and Hunyuan panoramic images. This benchmark also includes a highly challenging test set with 1,707 QA pairs, comprising not only a carefully selected subset from the training distribution but also newly added Sora-generated videos and Hunyuan panoramic images containing multiple safety hazards, used to evaluate the model's robustness in complex safety scenarios. Extensive experiments on 22 popular VLMs demonstrate that current VLMs lack robust capabilities for safety hazard assessment. Importantly, models trained on the TSHA training set achieve a significant performance improvement of up to +18.3 points on the TSHA test set and also exhibit enhanced generalizability across other benchmarks, underscoring the substantial contribution and importance of the TSHA benchmark.

Recent advances in Large Language Models (LLMs) and Text-to-Image (T2I) models have led to the emergence of Unified Multimodal Models (UMMs), where multimodal understanding and image generation are tightly integrated within a shared architecture. Prior studies suggest that such reciprocity enhances cross-functionality performance through shared representations and joint optimization. However, the safety implications of this tight coupling remain largely unexplored, as existing safety research predominantly analyzes understanding and generation functionalities in isolation. In this work, we investigate whether cross-functionality reciprocity itself constitutes a structural source of vulnerability in UMMs. We propose RICE: Reciprocal Interaction-based Cross-functionality Exploitation, a novel attack paradigm that explicitly exploits bidirectional interactions between understanding and generation. Using this framework, we systematically evaluate Generation-to-Understanding (G-U) and Understanding-to-Generation (U-G) attack pathways, demonstrating that unsafe intermediate signals can propagate across modalities and amplify safety risks. Extensive experiments show high Attack Success Rates (ASR) in both directions, revealing previously overlooked safety weaknesses inherent to UMMs.

Recent research points toward LLMs being manipulated through adversarial and seemingly benign inputs, resulting in harmful, biased, or policy-violating outputs. In this paper, we study an underexplored issue concerning harmful and toxic mathematical word problems. We show that math questions, particularly those framed as natural language narratives, can serve as a subtle medium for propagating biased, unethical, or psychologically harmful content, with heightened risks in educational settings involving children. To support a systematic study of this phenomenon, we introduce ToxicGSM, a dataset of 1.9k arithmetic problems in which harmful or sensitive context is embedded while preserving mathematically well-defined reasoning tasks. Using this dataset, we audit the behaviour of existing LLMs and analyse the trade-offs between safety enforcement and mathematical correctness. We further propose SafeMath -- a safety alignment technique that reduces harmful outputs while maintaining, and in some cases improving, mathematical reasoning performance. Our results highlight the importance of disentangling linguistic harm from math reasoning and demonstrate that effective safety alignment need not come at the cost of accuracy.

While prior red-teaming efforts have focused on eliciting harmful text outputs from large language models (LLMs), such approaches fail to capture agent-specific vulnerabilities that emerge through multi-step tool execution, particularly in rapidly growing ecosystems such as the Model Context Protocol (MCP). To address this gap, we propose a trajectory-aware evolutionary search method, T-MAP, which leverages execution trajectories to guide the discovery of adversarial prompts. Our approach enables the automatic generation of attacks that not only bypass safety guardrails but also reliably realize harmful objectives through actual tool interactions. Empirical evaluations across diverse MCP environments demonstrate that T-MAP substantially outperforms baselines in attack realization rate (ARR) and remains effective against frontier models, including GPT-5.2, Gemini-3-Pro, Qwen3.5, and GLM-5, thereby revealing previously underexplored vulnerabilities in autonomous LLM agents. Code is available at https://github.com/pwnhyo/T-MAP.

Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation, failing to capture whether a model is simultaneously pedagogically effective and safe across student-tutor interaction. We argue that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic content but the quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding. To systematically study this failure mode, we introduce SafeTutors, a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry. SafeTutors is organized around a theoretically grounded risk taxonomy comprising 11 harm dimensions and 48 sub-risks drawn from learning-science literature. We uncover that all models show broad harm; scale doesn't reliably help; and multi-turn dialogue worsens behavior, with pedagogical failures rising from 17.7% to 77.8%. Harms also vary by subject, so mitigations must be discipline-aware, and single-turn "safe/helpful" results can mask systematic tutor failure over extended interaction.

Large Vision-Language Models (LVLMs) have shown remarkable potential across a wide array of vision-language tasks, leading to their adoption in critical domains such as finance and healthcare. However, their growing deployment also introduces significant security and privacy risks. Malicious actors could potentially exploit these models to extract sensitive information, highlighting a critical vulnerability. Recent studies show that LVLMs often fail to consistently refuse instructions designed to compromise user privacy. While existing work on privacy protection has made meaningful progress in preventing the leakage of sensitive data, they are constrained by limitations in both generalization and non-destructiveness. They often struggle to robustly handle unseen privacy-related queries and may inadvertently degrade a model's performance on standard tasks. To address these challenges, we introduce Neural Gate, a novel method for mitigating privacy risks through neuron-level model editing. Our method improves a model's privacy safeguards by increasing its rate of refusal for privacy-related questions, crucially extending this protective behavior to novel sensitive queries not encountered during the editing process. Neural Gate operates by learning a feature vector to identify neurons associated with privacy-related concepts within the model's representation of a subject. This localization then precisely guides the update of model parameters. Through comprehensive experiments on MiniGPT and LLaVA, we demonstrate that our method significantly boosts the model's privacy protection while preserving its original utility. The code is available at https://github.com/Xiangkui-Cao/Neural-Gate.

Existing LLM safety evaluations rely on binary attack-success rates and domain-agnostic taxonomies, leaving regulated Banking, Financial Services, and Insurance (BFSI) deployments exposed to failures elicited through legally or professionally plausible framing. We introduce RAHS (Risk-Adjusted Harm Score), a risk-sensitive metric jointly capturing disclosure severity, disclaimer mitigation, and inter-judge agreement, and FinRedTeamBench, a 989-prompt benchmark spanning seven BFSI risk areas and 34 sub-categories mapped to regulatory frameworks. Evaluation uses an ensemble of three heterogeneous LLM judges, validated against human experts, and an adaptive multi-turn red-teaming pipeline. On nine open-weight models, RAHS preserves separation under near-ceiling ASR, ranking is stable under hyperparameter sweeps, and multi-turn pressure drives not only more jailbreaks but more operationally severe disclosures, exposing failure modes that single-turn, domain-agnostic evaluations cannot reveal.

With the global proliferation of Large Language Models (LLMs), cultural safety, defined as the ability to generate respectful and appropriate responses across diverse cultures, becomes critical for responsible AI deployment. However, existing research often treats cultural safety and cultural knowledge in isolation. It remains unclear whether cultural safety is grounded in understanding varying cultural knowledge to enable LLMs to adaptively yield respectful and appropriate responses across diverse cultures, which trigger ethical concerns in cross-cultural scenarios. In this work, we introduce AdaCultureSafe, a dataset designed to jointly evaluate cultural safety and knowledge. Through AdaCultureSafe, we reveal a critical insight: \textit{a significant decoupling exists between cultural safety and cultural knowledge proficiency. Although LLMs possess rich cultural knowledge, they fail to leverage it to improve cultural safety.} LLMs tend to rely on generic safety rather than safety based on culture-specific knowledge. Motivated by this, we propose a knowledge-grounded method that elicits internal cultural knowledge in LLMs during response generation. Experimental results demonstrate that our approach effectively improves cultural safety. Our work can aid better understanding the landscape of cultural safety of LLMs.

Large language model (LLM)-based judges are widely adopted for automated evaluation and reward modeling, yet their judgments are often affected by judgment biases. Accurately evaluating these biases is essential for ensuring the reliability of LLM-based judges. However, existing studies typically investigate limited biases under a single judge formulation, either generative or discriminative, lacking a comprehensive evaluation. To bridge this gap, we propose JudgeBiasBench, a benchmark for systematically quantifying biases in LLM-based judges. JudgeBiasBench defines a taxonomy of judgment biases across 4 dimensions, and constructs bias-augmented evaluation instances through a controlled bias injection pipeline, covering 12 representative bias types. We conduct extensive experiments across both generative and discriminative judges, revealing that current judges exhibit significant and diverse bias patterns that often compromise the reliability of automated evaluation. To mitigate judgment bias, we propose bias-aware training that explicitly incorporates bias-related attributes into the training process, encouraging judges to disentangle task-relevant quality from bias-correlated cues. By adopting reinforcement learning for generative judges and contrastive learning for discriminative judges, our methods effectively reduce judgment biases while largely preserving general evaluation capability.

Language models are commonly fine-tuned for safety alignment to refuse harmful prompts. One approach fine-tunes them to generate categorical refusal tokens that distinguish different refusal types before responding. In this work, we leverage a version of Llama 3 8B fine-tuned with these categorical refusal tokens to enable inference-time control over fine-grained refusal behavior, improving both safety and reliability. We show that refusal token fine-tuning induces separable, category-aligned directions in the residual stream, which we extract and use to construct categorical steering vectors with a lightweight probe that determines whether to steer toward or away from refusal during inference. In addition, we introduce a learned low-rank combination that mixes these category directions in a whitened, orthonormal steering basis, resulting in a single controllable intervention under activation-space anisotropy, and show that this intervention is transferable across same-architecture model variants without additional training. Across benchmarks, both categorical steering vectors and the low-rank combination consistently reduce over-refusals on benign prompts while increasing refusal rates on harmful prompts, highlighting their utility for multi-category refusal control.

Safety evaluation of multimodal large language models requires tracking not only whether an attack succeeds, but also how the interaction unfolds across turns and input modalities. We present MUSE (Multimodal Unified Safety Evaluation), an open-source, browser-based, run-centric platform for multimodal safety evaluation. MUSE treats each attack run as the persistent unit of execution, inspection, and analysis, preserving its configuration, multi-turn trajectory, delivered modalities and media, target responses, and safety judgments. A five-level response taxonomy further distinguishes full Compliance from Partial Compliance and refusal behavior, yielding hard ASR, soft ASR, and gray-zone width (GZW). Across 11,700 evaluations on six multimodal LLMs, direct text-only requests yield only 3.1% macro hard ASR and 4.4% soft ASR, while iterative attack procedures are substantially more effective. Attack effectiveness also varies substantially with the attacker backbone. In contrast, Inter-Turn Modality Switching (ITMS), evaluated as a controlled delivery-modality probe, does not consistently increase attack success. These results demonstrate the value of run-centric, fine-grained evaluation for characterizing multimodal safety behavior beyond a single binary success metric.

LLMs see the world as a single stream of text, partitioned into roles like <user> or <tool>. We trace prompt injection to role confusion: models perceive the source of text from how it sounds, not its labeled role. A command hidden in a webpage hijacks an agent simply because it sounds like <user> text, despite its <tool> label. We design role probes to measure how LLMs internally perceive "who is speaking," and find that injected text occupies the same representational space as the trusted role it imitates. We demonstrate this with CoT Forgery, a zero-shot attack that injects fabricated reasoning into user prompts and tool outputs. Models mistake the forgery for their own thoughts, yielding 60% attack success against frontier models with near-zero baselines. Strikingly, the degree of role confusion predicts attack success before a single token is generated. This mechanism generalizes beyond CoT Forgery to standard agent prompt injections, revealing prompt injection as a measurable consequence of role perception. To the model, sounding like a role is indistinguishable from being one. Project page and writeup: https://role-confusion.github.io

Jailbreak attacks remain a persistent threat to large language model safety. We propose Context-Conditioned Delta Steering (CC-Delta), an SAE-based defense that identifies jailbreak-relevant sparse features by comparing token-level representations of the same harmful request with and without jailbreak context. Using paired harmful/jailbreak prompts, CC-Delta selects features via statistical testing and applies inference-time mean-shift steering in SAE latent space. Across four aligned instruction-tuned models and twelve jailbreak attacks, CC-Delta achieves comparable or better safety-utility tradeoffs than baseline defenses operating in dense latent space. In particular, our method clearly outperforms dense mean-shift steering on all four models, and particularly against out-of-distribution attacks, showing that steering in sparse SAE feature space offers advantages over steering in dense activation space for jailbreak mitigation. Our results suggest off-the-shelf SAEs trained for interpretability can be repurposed as practical jailbreak defenses without task-specific training.

Vision-Language Models (VLMs) are now a core part of modern AI. Recent work proposed several visual jailbreak attacks using single/ holistic images. However, contemporary VLMs demonstrate strong robustness against such attacks due to extensive safety alignment through preference optimization, e.g., reinforcement learning from human feedback (RLHF). In this work, we identify a new vulnerability: while VLM pretraining and instruction tuning generalize well to split-image inputs, safety alignment is typically performed only on holistic images and does not account for harmful semantics distributed across multiple image fragments. Consequently, VLMs often fail to detect and reject harmful split-image inputs, in which unsafe cues emerge only upon combining images. We introduce novel split-image visual jailbreak attacks (\textbf{SIVA}) that exploit this misalignment. Unlike prior optimization-based attacks, which exhibit poor black-box transferability due to architectural and prior mismatches across models, our attacks evolve in progressive phases from naive splitting to an adaptive white-box attack, culminating in a black-box transfer attack. Our strongest strategy leverages a novel adversarial knowledge distillation \textbf{(Adv-KD)} algorithm to substantially improve cross-model transferability. Evaluations on four state-of-the-art modern VLMs and three jailbreak datasets demonstrate that our strongest attack achieves up to 44% higher transfer success than existing baselines. Lastly, we propose efficient ways to address this critical vulnerability in the current VLM safety alignment.

Large language models can recover mid-generation from task-misaligned activation steering, producing explicit verbal restarts (e.g., ``wait, that's not right'') and continuing on-topic even while the steering perturbation remains active. We term this Endogenous Steering Resistance (ESR). Using sparse autoencoder (SAE) latents to steer model activations, we find that Llama-3.3-70B exhibits explicit ESR at 3.8%, with smaller models from the Llama-3 and Gemma-2 families showing the explicit form less frequently. Two controls dissociate ESR into a detection event and a sustained-resistance component that conditioning on recent on-topic tokens does not fully explain. We identify 26 SAE latents through contrastive on-topic/off-topic search; zero-ablating them reduces the multi-attempt rate by 25%, with random-latent and held-out-prompt controls supporting specificity. ESR can also be deliberately enhanced through both meta-prompting and fine-tuning on synthetic self-correction examples. ESR has dual implications for safety: it could harden models against adversarial activation-space manipulation, but may equally interfere with beneficial steering-based interventions, since the model has no way to distinguish the two. Code is available at https://github.com/agencyenterprise/endogenous-steering-resistance.

From generating headlines to fabricating news, the Large Language Models (LLMs) are typically assessed by their final outputs, under the safety assumption that a refusal response signifies safe reasoning throughout the entire process. Challenging this assumption, our study reveals that during fake news generation, even when a model rejects a harmful request, its Chain-of-Thought (CoT) reasoning may still internally contain and propagate unsafe narratives. To analyze this phenomenon, we introduce a unified safety-analysis framework that systematically deconstructs CoT generation across model layers and evaluates the role of individual attention heads through Jacobian-based spectral metrics. Within this framework, we introduce three interpretable measures: stability, geometry, and energy to quantify how specific attention heads respond or embed deceptive reasoning patterns. Extensive experiments on multiple reasoning-oriented LLMs show that the generation risk rises significantly when the thinking mode is activated, where the critical routing decisions are concentrated in only a few contiguous mid-depth layers. By precisely identifying the attention heads responsible for this divergence, our work challenges the assumption that refusal implies safety and provides a new understanding perspective for mitigating latent reasoning risks.

Modern deployments require LLMs to enforce safety policies at scale, yet many controls rely on inference-time interventions that add recurring compute cost and serving complexity. Activation steering is widely used, but it requires runtime hooks and scales cost with the number of generations; conditional variants improve selectivity by gating when steering is applied but still retain an inference-time control path. We ask whether selective refusal can be moved entirely offline: can a mechanistic understanding of category-specific refusal be distilled into a circuit-restricted weight update that deploys as a standard checkpoint? We propose C-Δθ Circuit Restricted Weight Arithmetic}, which (i) localizes refusal-causal computation as a sparse circuit using EAP-IG and (ii) computes a constrained weight update ΔθC supported only on that circuit (typically <5% of parameters). Applying ΔθC yields a drop-in edited checkpoint with no inference-time hooks, shifting cost from per request intervention to a one-time offline update. We evaluate category-targeted selectivity and capability retention on refusal and utility benchmarks.

Hallucinations remain a major obstacle for large language models (LLMs), especially in safety-critical domains. We present HALT (Hallucination Assessment via Log-probs as Time series), a lightweight hallucination detector that leverages only the top-20 token log-probabilities from LLM generations as a time series. HALT uses a gated recurrent unit model combined with entropy-based features to learn model calibration bias, providing an extremely efficient alternative to large encoders. Unlike white-box approaches, HALT does not require access to hidden states or attention maps, relying only on output log-probabilities. Unlike black-box approaches, it operates on log-probs rather than surface-form text, which enables stronger domain generalization and compatibility with proprietary LLMs without requiring access to internal weights. To benchmark performance, we introduce HUB (Hallucination detection Unified Benchmark), which consolidates prior datasets into ten capabilities covering both reasoning tasks (Algorithmic, Commonsense, Mathematical, Symbolic, Code Generation) and general purpose skills (Chat, Data-to-Text, Question Answering, Summarization, World Knowledge). While being 30x smaller, HALT outperforms Lettuce, a fine-tuned modernBERT-base encoder, achieving a 60x speedup gain on HUB. HALT and HUB together establish an effective framework for hallucination detection across diverse LLM capabilities.

Large Language Models (LLMs) are increasingly deployed in tool-driven environments such as healthcare analytics, financial systems, retrieval-augmented generation (RAG), and multi-agent workflows. Although the Model Context Protocol (MCP) standardizes how LLM applications expose and invoke external tools, its baseline model does not require tool-use manifests to be cryptographically authenticated, freshness-checked, policy-bound, or independently auditable before execution. As a result, MCP pipelines may remain vulnerable to manifest tampering, unauthorized tool invocation, replay of stale requests, and weak accountability. This paper presents a manifest-level enforcement layer for MCP-based LLM pipelines. It treats each MCP tool-use manifest as a first-class security object whose canonical form must be policy-validated, freshness-checked, digitally signed, verified before execution, and linked to tamper-evident audit evidence. The framework binds tool invocation to verifiable manifest integrity and fail-closed authorization, separates user-visible request parameters from execution metadata, rejects non-compliant or stale manifests before execution, and records accepted invocations in a Merkle-based transparency log. Evaluation across GPT-5.3, LLaMA-3.5, and DeepSeek-V3 using up to 50,000 manifest instances shows near-linear scalability (R^2 = 0.998), bounded verification latency (<= 9.4 ms on edge devices), and rejection of expired, malformed, replayed, and policy-violating manifests, with rejection rates above 98.7%. Experiments in healthcare, finance, RAG, and multi-agent settings show that manifest-level cryptographic enforcement enables low-overhead, traceable, and auditable execution control for heterogeneous LLM-tool pipelines.

Semantic caching has emerged as a pivotal technique for scaling LLM applications, widely adopted by major providers including AWS and Microsoft. By utilizing semantic embedding vectors as cache keys, this mechanism effectively minimizes latency and redundant computation for semantically similar queries. In this work, we conceptualize semantic cache keys as a form of fuzzy hashes. We demonstrate that the locality required to maximize cache hit rates fundamentally conflicts with the cryptographic avalanche effect necessary for collision resistance. Our conceptual analysis formalizes this inherent trade-off between performance (locality) and security (collision resilience), revealing that semantic caching is naturally vulnerable to key collision attacks. While prior research has focused on side-channel and privacy risks, we present the first systematic study of integrity risks arising from cache collisions. We introduce CacheAttack, an automated framework for launching black-box collision attacks. We evaluate CacheAttack in security-critical tasks and agentic workflows. It achieves a hit rate of 86% in LLM response hijacking and can induce malicious behaviors in LLM agent, while preserving strong transferability across different embedding models. A case study on a financial agent further illustrates the real-world impact of these vulnerabilities. Finally, we discuss mitigation strategies.

Modern LLMs are increasingly accessed via black-box APIs, requiring users to transmit sensitive prompts, outputs, and fine-tuning data to external providers, creating a critical privacy risk at the API boundary. We introduce AlienLM, a deployable API-only \cradd{exposure-reduction layer that reduces plaintext exposure} by translating text into an Alien Language via a vocabulary-scale bijection, enabling lossless recovery on the client side. Using only standard fine-tuning APIs, Alien Adaptation Training (AAT) adapts target models to operate directly on alienized inputs. Across four LLM backbones and seven benchmarks, AlienLM retains over 81% of plaintext-oracle performance on average, substantially outperforming random-bijection and character-level baselines. Under adversaries with access to model weights, corpus statistics, and learning-based inverse translation, recovery attacks reconstruct fewer than 0.22% of alienized tokens. Our results demonstrate a practical pathway for \cradd{privacy-aware} LLM deployment under API-only access, substantially reducing plaintext exposure while maintaining task performance. Code and data are available at https://github.com/KimJaehee0725/AlienLM.

Autonomous code agents built on large language models are reshaping software and AI development through tool use, long-horizon reasoning, and self-directed interaction. However, this autonomy introduces a previously unrecognized security risk: agentic interaction fundamentally expands the LLM attack surface, enabling systematic probing and recovery of hidden system prompts that guide model behavior. We identify system prompt extraction as an emergent vulnerability intrinsic to code agents and present \textbf{\textsc{JustAsk}}, a self-evolving framework that autonomously discovers effective extraction strategies through interaction alone. Unlike prior prompt-engineering or dataset-based attacks, \textsc{JustAsk} requires no handcrafted prompts, labeled supervision, or privileged access beyond standard user interaction. It formulates extraction as an online exploration problem, using Upper Confidence Bound-based strategy selection and a hierarchical skill space spanning atomic probes and high-level orchestration. These skills exploit imperfect system-instruction generalization and inherent tensions between helpfulness and safety. Evaluated on \textbf{41} black-box commercial models across multiple providers, \textsc{JustAsk} consistently achieves full or near-complete system prompt recovery, revealing recurring design- and architecture-level vulnerabilities. Our results expose system prompts as a critical yet largely unprotected attack surface in modern agent systems.

Safety alignment in Large Language Models is critical for healthcare; however, reliance on binary refusal boundaries often results in over-refusal of benign queries or unsafe compliance with harmful ones. While existing benchmarks measure these extremes, they fail to evaluate Safe Completion: the model's ability to maximise helpfulness on dual-use or borderline queries by providing safe, high-level guidance without crossing into actionable harm. We introduce Health-ORSC-Bench, the first large-scale benchmark designed to systematically measure Over-Refusal and Safe Completion quality in healthcare. Comprising 31,920 benign boundary prompts across seven health categories (e.g., self-harm, medical misinformation), our framework uses an automated pipeline with human validation to test models at varying levels of intent ambiguity. We evaluate 30 state-of-the-art LLMs, including GPT-5 and Claude-4, revealing a significant tension: safety-optimised models frequently refuse up to 80% of "Hard" benign prompts, while domain-specific models often sacrifice safety for utility. Our findings demonstrate that model family and size significantly influence calibration: larger frontier models (e.g., GPT-5, Llama-4) exhibit "safety-pessimism" and higher over-refusal than smaller or MoE-based counterparts (e.g., Qwen-3-Next), highlighting that current LLMs struggle to balance refusal and compliance. Health-ORSC-Bench provides a rigorous standard for calibrating the next generation of medical AI assistants toward nuanced, safe, and helpful completions. Furthermore, our benchmark facilitates reproducible evaluation, encourages safety calibration, and supports development of clinically reliable, context-aware, human-aligned medical AI systems. Our code and data are available at: https://github.com/ZhihaoZhang97/Health-ORSC-Bench. Warning: Some contents may include toxic or undesired contents.

As large language models (LLMs) are increasingly deployed in real-world applications, safety guardrails are required to go beyond coarse-grained filtering and support fine-grained, interpretable, and adaptable risk assessment. However, existing solutions often rely on rapid classification schemes or post-hoc rules, resulting in limited transparency, inflexible policies, or prohibitive inference costs. To this end, we present YuFeng-XGuard, a reasoning-centric guardrail model family designed to perform multi-dimensional risk perception for LLM interactions. Instead of producing opaque binary judgments, YuFeng-XGuard generates structured risk predictions, including explicit risk categories and configurable confidence scores, accompanied by natural language explanations that expose the underlying reasoning process. This formulation enables safety decisions that are both actionable and interpretable. To balance decision latency and explanatory depth, we adopt a tiered inference paradigm that performs an initial risk decision based on the first decoded token, while preserving ondemand explanatory reasoning when required. In addition, we introduce a dynamic policy mechanism that decouples risk perception from policy enforcement, allowing safety policies to be adjusted without model retraining. Extensive experiments on a diverse set of public safety benchmarks demonstrate that YuFeng-XGuard achieves stateof-the-art performance while maintaining strong efficiency-efficacy trade-offs. We release YuFeng-XGuard as an open model family, including both a full-capacity variant and a lightweight version, to support a wide range of deployment scenarios.

Large language models (LLMs) are increasingly used for hardware and firmware code generation, but existing studies primarily evaluate functional correctness while largely overlooking security. However, LLM-generated code that appears functionally sound may embed security flaws which could induce catastrophic damages after deployment. This critical research gap motivates us to design a benchmark for assessing security awareness under realistic specifications. In this work, we introduce HardSecBench, a benchmark with 924 tasks spanning Verilog Register Transfer Level (RTL) and firmware-level C, covering 76 hardware-relevant Common Weakness Enumeration (CWE) entries. Each task includes a structured specification, a secure reference implementation, and executable tests. To automate artifact synthesis, we propose a multi-agent pipeline that decouples synthesis from verification and grounds evaluation in execution evidence, enabling reliable evaluation. We evaluate diverse LLMs and find that they often satisfy functional requirements while leaving security risks. We also find that security results vary with prompting. These findings highlight pressing challenges and offer actionable insights for future advancements in LLM-assisted hardware design. Our data and code are available at https://github.com/chenqirui2002/HardSecBench.

Benchmarking large language models (LLMs) is critical for understanding their capabilities, limitations, and robustness. In addition to interface artifacts, prior studies have shown that LLM decisions can be influenced by directive signals such as social cues, framing, and instructions. In this work, we introduce option injection, a benchmarking approach that augments the multiple-choice question answering (MCQA) interface with an additional option containing a misleading directive, leveraging standardized choice structure and scalable evaluation. We construct OI-Bench, a benchmark of 3,000 questions spanning knowledge, reasoning, and commonsense tasks, with 16 directive types covering social compliance, bonus framing, threat framing, and instructional interference. This setting combines manipulation of the choice interface with directive-based interference, enabling systematic assessment of model susceptibility. We evaluate 12 LLMs to analyze attack success rates, behavioral responses, and further investigate mitigation strategies ranging from inference-time prompting to post-training alignment. Experimental results reveal substantial vulnerabilities and heterogeneous robustness across models. OI-Bench is expected to support more systematic evaluation of LLM robustness to directive interference within choice-based interfaces.

Quantization is widely adopted to reduce the computational cost of large language models (LLMs); however, its implications for fairness and safety, particularly in dynamic quantization and multilingual contexts, remain underexplored. In this work, we conduct a systematic study of how static and dynamic quantization methods impact fairness and safety across benchmarks measuring intrinsic and extrinsic bias and safety alignment. For fairness, we evaluate English, French, Dutch, Spanish, and Turkish; for safety, we focus on English, Korean, and Arabic. Our findings reveal that quantization consistently degrades fairness and safety, with dynamic methods demonstrating greater stability than static ones. Moreover, fairness degradation varies across languages, while safety deterioration is especially pronounced in non-English settings. To address these risks, we introduce Critical Weight Protection, a novel technique that identifies and preserves fairness- and safety-critical weights during quantization. This approach effectively mitigates bias and safety deterioration without costly retraining or alignment, maintaining trustworthiness while retaining efficiency.

Large language models can produce toxic or inappropriate text even for benign inputs, creating risks when deployed at scale. Detoxification is therefore important for safety and user trust, particularly when we want to reduce harmful content without sacrificing the model's generation quality. Many existing approaches rely on model retraining, gradients, or learned auxiliary components, which can be costly and may not transfer across model families or to truly black-box settings. We introduce a test-time procedure that approximates the gradient of completion toxicity with respect to the input embeddings and uses a small number of descent steps to steer generation toward less toxic continuations. This is achieved with zeroth-order optimization that requires only access to input embeddings, a toxicity scoring function, and forward evaluations of the model. Empirically, the approach delivers robust toxicity reductions across models and prompts and, in most settings, achieves the best overall toxicity-quality trade-off. More broadly, our work positions word embeddings as effective control variables and encourages wider use of black-box optimization to guide autoregressive language models toward scalable, safer text generation, without requiring any training or access to intermediate computations.

Recent results in learning a language in the limit have shown that, although language identification is impossible, language generation is tractable. As this foundational area expands, we need to consider the implications of language generation in real-world settings. This work offers the first theoretical treatment of safe language generation. Building on the computational paradigm of learning in the limit, we formalize the tasks of safe language identification and generation. We prove that under this model, safe language identification is impossible, and that safe language generation is at least as hard as (vanilla) language identification, which is also impossible. Last, we discuss several intractable and tractable cases.

Current safety evaluations of large language models (LLMs) create a dangerous illusion of universal protection by aggregating harms under generic categories such as "Identity Hate", obscuring vulnerabilities toward specific populations. In this work, we expose the Selective Safety Trap: a systemic failure mode where models robustly defend specific populations while leaving underrepresented communities highly vulnerable to identical adversarial attacks. To systematically audit this phenomenon, we introduce MiJaBench, a bilingual (English-Portuguese) adversarial benchmark comprising 43,961 controlled jailbreaking prompts across 16 minority groups. By evaluating 14 state-of-the-art LLMs on MiJaBench, we curate 615,454 prompt-response pairs that compose MiJaBench-Align, revealing that safety alignment is not a uniform semantic capability but a demographic hierarchy, with defense rates fluctuating by up to 42% within the same model solely based on the target group. This disparity persists across architectures and languages and is amplified by scaling, indicating that current alignment methods learn group-specific safeguards rather than a generalized notion of harm. Through targeted direct preference optimization (DPO) on a 1B-parameter baseline, we achieve strong zero-shot safety generalizations to entirely unseen demographics and complex attack strategies. We release all datasets and scripts to provide the community with a concrete pathway toward equitable, transferable safety alignment.

State space models (SSMs) like Mamba offer efficient alternatives to Transformer-based language models, with linear time complexity. Yet, their adversarial robustness remains critically unexplored. This paper studies the phenomenon whereby specific short input phrases induce a partial amnesia effect in such models, by irreversibly overwriting information in their hidden states, referred to as a Hidden State Poisoning Attack (HiSPA). Our benchmark RoBench-25 allows evaluating a model's information retrieval capabilities when subject to HiSPAs, and confirms the vulnerability of SSMs against such attacks. Even the recent Jamba-1.7-Mini SSM--Transformer (a 52B hybrid model) collapses on RoBench-25 under some HiSPA triggers, whereas pure Transformers do not. We also observe that HiSPA triggers significantly weaken the Jamba model on the popular Open-Prompt-Injections benchmark, unlike pure Transformers. We further show that the theoretical and empirical findings extend to Mamba-2, and also analyse a Mamba-2-based hybrid (Nemotron-3-Nano). Finally, our interpretability study reveals patterns in Mamba's hidden layers during HiSPAs that could be used to build a HiSPA mitigation system. The full code and data to reproduce the experiments can be found at https://github.com/TortueSagace/hispa.

Generative AI models reproduce the human biases in their training data and further amplify them through mechanisms such as mode collapse. The loss of diversity produces homogenization, which not only harms the minoritized but impoverishes everyone. We argue homogenization should be a central concern in AI safety. To meaningfully characterize homogenization in Large Language Models (LLMs), we introduce a framework that allows stakeholders to encode their context and value system. We illustrate our approach with an experiment that surfaces gender bias in an LLM (Claude 3.5 Haiku) on an open-ended story prompt. Building from queer theory, we formalize homogenization in terms of normativity. Borrowing language from feminist theory, we introduce the concept of xeno-reproduction as a class of tasks for mitigating homogenization by promoting diversity. Our work opens a collaborative line of research that seeks to understand and advance diversity in AI.

AI coding assistants produce vulnerable code in 45% of security-relevant scenarios~\cite{veracode2025}, yet no public training dataset teaches both traditional web security and AI/ML-specific defenses in a format suitable for instruction tuning. We present SecureCode, a production-grade dataset of 2,185 multi-turn security training examples spanning two domains: web application security (1,435 examples covering the OWASP Top 10 2021 across 11 languages and 9 frameworks, 100% grounded in documented CVEs and security incidents) and AI/ML security (750 examples covering all 10 OWASP LLM Top 10 2025 categories across more than 40 frameworks, including LangChain, OpenAI, and Hugging Face). Every example follows a 4-turn conversational structure -- feature request; vulnerable and secure implementations with attack demonstrations; advanced probing; and defense-in-depth operational guidance -- designed for direct use in instruction tuning pipelines. Quality assurance combines automated structural validation with multi-agent review from seven specialist AI perspectives (more than 10{,}500 assessments) and an 8-phase remediation pipeline, producing a rubric-calibrated mean quality score of 93.8/100 ($σ= 0.93$) for the AI/ML component. Each example provides SIEM integration strategies, infrastructure hardening recommendations, and testing approaches using production frameworks. We release the unified dataset on Hugging Face with domain-specific loading configurations (web, aiml, default), alongside eight fine-tuned open-source models (3B--20B parameters, QLoRA), and an evaluation framework with four security-specific metrics. To our knowledge, SecureCode is the first public dataset that jointly provides OWASP Top 10 2021 web coverage and OWASP LLM Top 10 2025 AI/ML coverage in a unified conversational schema suitable for instruction tuning.

Disclaimer: Samples in this paper may be harmful and cause discomfort. Multimodal large language models (MLLMs) enable multimodal understanding but inherit toxic signals from weakly curated pretraining corpora, leading to explicitly toxic outputs, especially under adversarial triggers that late, opaque training-free detoxification methods struggle to handle. We propose SGM, a white-box neuron-level multimodal intervention that acts like safety glasses for toxic neurons: it recalibrates a set of toxicity-associated neurons via expertise-weighted soft suppression, neutralizing harmful cross-modal activations without any parameter updates. We establish MM-TOXIC-QA, a multimodal toxicity data framework, and compare SGM with existing detoxification techniques. Experiments on open-source MLLMs show that SGM mitigates explicit toxicity in standard and adversarial conditions, cutting average harmful rates from 45.0% to 4.5% while preserving fluency and multimodal reasoning. SGM is extensible, and its combined defenses, denoted as SGM*, integrate with existing detoxification methods for stronger safety performance, providing an interpretable, low-cost solution for toxicity-controlled multimodal generation.

As organizations increasingly deploy LLMs in sensitive domains such as legal, financial, and medical settings, ensuring alignment with internal organizational policies has become a priority. Existing content moderation frameworks remain largely confined to the safety domain and lack the robustness to capture nuanced organizational policies. LLM-as-a-judge and fine-tuning approaches, though flexible, introduce significant latency and training cost. To address these limitations, we frame policy violation detection as an out-of-distribution (OOD) problem in the model's activation space. We propose a training-free method that operates directly on the LLM internal representations, leveraging prior evidence that decision-relevant information is encoded within them. Inspired by whitening techniques, we derive policy-violation scores directly from normalized representations of LLM hidden activations. Our method requires only the policy text and a small number of illustrative samples, making it lightweight and easily deployable. We extensively evaluate our method across multiple LLMs and challenging policy benchmarks. It achieves up to 86.0% F1, outperforming both fine-tuned and LLM-as-a-judge baselines while requiring substantially less computation. Our code is publicly available at: https://github.com/FujitsuResearch/LLM-policy-violation-detection

Vibe coding is a new software development paradigm in which human engineers prompt a large language model (LLM) agent to complete complex coding tasks with little supervision. Although vibe coding is increasingly adopted, is the generated code really safe to deploy in production? To investigate this question, we propose SUSVIBES, a benchmark consisting of 186 feature-request software engineering tasks from real-world open-source projects, for which, human programmers committed vulnerable implementations. We evaluate 12 widely used coding agentic settings with frontier models on the benchmark. Disturbingly, all agents perform poorly in terms of software security. Although 57% of the solutions from SWE-Agent with Claude 4 Sonnet are functionally correct, only 11.8% are secure. Further experiments demonstrate that preliminary security strategies, such as augmenting the feature request with vulnerability hints, cannot mitigate these security issues. Our findings raise serious concerns about the widespread adoption of vibe coding, particularly in security-sensitive applications. The code and dataset are available at https://github.com/LeiLiLab/susvibes. The leaderboard is at https://leililab.github.io/susvibes-leaderboard.

When answering user queries, LLMs often retrieve knowledge from external sources stored in retrieval-augmented generation (RAG) databases. These are often populated from unvetted sources, e.g. the open web, and can contain maliciously crafted data. This paper studies attacks that can manipulate the context retrieved by LLMs from such RAG databases. Prior work on such context manipulation primarily injects false or toxic content, which can often be detected by fact-checking or linguistic analysis. A more subtle threat, which we call epistemic bias injection (EBI), is where adversaries inject factually correct yet epistemically biased passages that systematically favor one side of an open-ended issue. Although linguistically coherent and truthful, such adversarial passages effectively crowd out alternative viewpoints during retrieval from the RAG and push LLM outputs towards an attack-desired stance. As a core contribution, we propose a novel characterization of the problem: We give a geometric metric that quantifies stance polarity and epistemic bias. This metric can be computed directly on embeddings of text passages. Leveraging it, we construct EBI attacks and develop a lightweight prototype defense called BiasDef for them. We evaluate them both on a comprehensive benchmark constructed from public question answering datasets. Our results show that: (1) the proposed attack induces significant stance polarity shifts, effectively evading existing retrieval-based sanitization defenses, and (2) BiasDef substantially reduces adversarial retrieval and epistemic bias in LLM's answers. Overall, this demonstrates the new threat as well as the ease of employing epistemic bias metrics for filtering in RAG-enabled LLMs.

As Large Language Model (LLM) agents increasingly operate in complex environments with real-world consequences, their safety becomes critical. While uncertainty quantification is well-studied for single-turn tasks, multi-turn agentic scenarios with real-world tool access present unique challenges where uncertainties and ambiguities compound, leading to severe or catastrophic risks beyond traditional text generation failures. We propose using "quitting" as a simple yet effective behavioral mechanism for LLM agents to recognize and withdraw from situations where they lack confidence. Leveraging the ToolEmu framework, we conduct a systematic evaluation of quitting behavior across 12 state-of-the-art LLMs. Our results demonstrate a highly favorable safety-helpfulness trade-off: agents prompted to quit with explicit instructions improve safety by an average of +0.39 on a 0-3 scale across all models (+0.64 for proprietary models), while maintaining a negligible average decrease of -0.03 in helpfulness. Our analysis demonstrates that simply adding explicit quit instructions proves to be a highly effective safety mechanism that can immediately be deployed in existing agent systems, and establishes quitting as an effective first-line defense mechanism for autonomous agents in high-stakes applications.

Large Language Models (LLMs) are increasingly used as interfaces to information, code, and real-world services, making prompt-level security failures a practical concern. Although jailbreak attacks, defenses, datasets, and automated judgers have advanced rapidly, evaluation remains fragmented across threat models, access assumptions, cost budgets, datasets, and success criteria. This makes reported attack success rates and defense gains hard to compare. This SoK systematizes LLM prompt security across concepts, data, tooling, and measurement. We propose linked taxonomies for jailbreak attacks, defenses, and model vulnerabilities, while separating technical mechanisms from attacker and defender capabilities. We also formalize threat, access, and cost assumptions as explicit evaluation metadata. To support reproducible evaluation, we release JailbreakDB, PromptSecurity-Eval, and PromptSecurity, a modular platform that represents each experiment as a tuple of model, attack, defense, dataset, and judger. Using matched evaluations across models, attacks, defenses, and judgers, we show that access regime, native harmful-query behavior, attack cost, defense backfire, taxonomy subcategory, and judger choice all materially affect security conclusions. Together, these artifacts support reproducible, cost-aware, and taxonomy-grounded evaluation of LLM prompt security. Leaderboard: https://datasec-lab.github.io/PromptSecurityLeaderboard/. Dataset: https://huggingface.co/datasets/youbin2014/JailbreakDB. GitHub: https://github.com/datasec-lab/PromptSecurity.

Large Language Model (LLM) agents are powering a growing share of interactive web applications, yet remain vulnerable to misuse and harm. Prior jailbreak research has largely focused on single-turn prompts, whereas real harassment often unfolds over multi-turn interactions. In this work, we present the Online Harassment Agentic Benchmark consisting of: (i) a synthetic multi-turn harassment conversation dataset, (ii) a multi-agent (e.g., harasser, victim) simulation informed by repeated game theory, (iii) three jailbreak methods attacking agents across memory, planning, and fine-tuning, and (iv) a mixed-methods evaluation framework. We utilize two prominent LLMs, LLaMA-3.1-8B-Instruct (open-source) and Gemini-2.0-flash (closed-source). Our results show that jailbreak tuning makes harassment nearly guaranteed with an attack success rate of 95.78--96.89% vs. 57.25--64.19% without tuning in Llama, and 99.33% vs. 98.46% without tuning in Gemini, while sharply reducing refusal rate to 1-2% in both models. The most prevalent toxic behaviors are Insult with 84.9--87.8% vs. 44.2--50.8% without tuning, and Flaming with 81.2--85.1% vs. 31.5--38.8% without tuning, indicating weaker guardrails compared to sensitive categories such as sexual or racial harassment. Qualitative evaluation further reveals that attacked agents reproduce human-like aggression profiles, such as Machiavellian/psychopathic patterns under planning, and narcissistic tendencies with memory. Counterintuitively, closed-source and open-source models exhibit distinct escalation trajectories across turns, with closed-source models showing significant vulnerability. Overall, our findings show that multi-turn and theory-grounded attacks not only succeed at high rates but also mimic human-like harassment dynamics, motivating the development of robust safety guardrails to ultimately keep online platforms safe and responsible.

Large language models (LLMs) are transforming everyday applications, yet deployment in cybersecurity lags due to a lack of high-quality, domain-specific models and training datasets. To address this gap, we present CyberPal 2.0, a family of cybersecurity-expert small language models (SLMs) ranging from 4B-20B parameters. To train CyberPal 2.0, we generate an enriched chain-of-thought cybersecurity instruction dataset built with our data enrichment and formatting pipeline, SecKnowledge 2.0, which integrates expert-in-the-loop steering of reasoning formats alongside LLM-driven multi-step grounding, yielding higher-fidelity, task-grounded reasoning traces for security tasks. Across diverse cybersecurity benchmarks, CyberPal 2.0 consistently outperforms its baselines and matches or surpasses various open and closed-source frontier models, while remaining a fraction of their size. On core cyber threat intelligence knowledge tasks, our models outperform almost all tested frontier models, ranking second only to Sec-Gemini v1. On core threat-investigation tasks, such as correlating vulnerabilities and bug tickets with weaknesses, our best 20B-parameter model outperforms GPT-4o, o1, o3-mini, and Sec-Gemini v1, ranking first, while our smallest 4B-parameter model ranks second.

The proliferation of Large Language Models (LLMs) has enabled a new class of psychologically grounded malicious comments, shifting fake news attacks from surface-level textual noise to deep cognitive and logical manipulation. This shift severely undermines existing detectors, which conventionally rely on static attack assumptions and fixed training distributions. To bridge this gap, we introduce CogniDir, an adaptive distributional learning framework that reformulates robust detection as a dynamic data mixture optimization problem for social media content safety. Grounded in cognitive psychology, we first formalize mechanism-specific cognitive adversarial paradigms to systematically expose deep-seated detector vulnerabilities. To address the vulnerability heterogeneity, CogniDir derives an information-theoretic score coupling empirical accuracy with probabilistic confidence, which is then mapped to adaptive sampling proportions through a Dirichlet-mean parameterization. This formulation enables smooth, feedback-driven reallocation of training exposure toward the most brittle attack mechanisms. Experimental results on three benchmarks demonstrate that CogniDir yields state-of-the-art robustness, improving F1 scores by up to 17.9% over competitive baselines under heterogeneous, AI-generated adversarial pressures.

The rapid proliferation of large language models (LLMs) has intensified the requirement for reliable safety evaluation to uncover model vulnerabilities. To this end, numerous LLM safety evaluation benchmarks are proposed. However, existing benchmarks generally rely on labor-intensive manual curation, which causes excessive time and resource consumption. They also exhibit significant redundancy and limited difficulty. To alleviate these problems, we introduce SafetyFlow, the first agent-flow system designed to automate the construction of LLM safety benchmarks. SafetyFlow can automatically build a comprehensive safety benchmark in only four days without any human intervention by orchestrating seven specialized agents, significantly reducing time and resource cost. Equipped with versatile tools, the agents of SafetyFlow ensure process and cost controllability while integrating human expertise into the automatic pipeline. The final constructed dataset, SafetyFlowBench, contains 23,446 queries with low redundancy and strong discriminative power. Our contribution includes the first fully automated benchmarking pipeline and a comprehensive safety benchmark. We evaluate the safety of 49 advanced LLMs on our dataset and conduct extensive experiments to validate our efficacy and efficiency.

We evaluate artificial intelligence (AI) systems without ground truth by exploiting a link between strategic gaming and information loss. Building on established information theory, we analyze which mechanisms resist adversarial manipulation. This motivates mutual evaluation, where the overseer is treated as a strategic player estimating mutual information by prompting, making truthful agent reporting an optimal strategy. We show that certain f-divergences, such as total variation distance (TVD), maintain polynomial guarantees under attack, building on an established exponential barrier for estimating mutual information (MI) in worst-case certification settings. Under adversarial attacks, TVD-MI maintains effectiveness (area under the curve 0.70--0.77) while other approaches can decay toward chance, demonstrating that prompting the same system for information relationships rather than quality judgments can improve robustness. The mechanisms decompose pairwise evaluations into reliable item-level detection scores without ground truth, addressing a key limitation of standard peer prediction. Pre-registration: https://osf.io/c7pum .

Large Language Models (LLMs), now a foundation in advancing natural language processing, power applications such as text generation, machine translation, and conversational systems. Despite their transformative potential, these models inherently rely on massive amounts of training data, often collected from diverse and uncurated sources, which exposes them to serious data security risks. Harmful or malicious data can compromise model behavior, leading to toxic outputs or hallucinations, while also creating vulnerabilities to data-driven attacks such as prompt injection and data poisoning. As LLMs continue to be integrated into critical real-world systems, understanding and addressing these data-centric security risks is imperative to safeguard user trust and system reliability. This survey offers a comprehensive overview of the main data security risks facing LLMs and reviews current defense strategies, including adversarial training, data cleaning, output guardrails, Reinforcement Learning from Human Feedback (RLHF), data augmentation, and Retrieval-Augmented Generation (RAG)/agent defenses. Additionally, we categorize and analyze relevant datasets used for assessing robustness and security across different domains, providing guidance for future research. Finally, we highlight key research directions that focus on data provenance and traceability, verifiable machine forgetting, secure model updates, standardized evaluation framework, explainability-driven security analysis, and effective governance frameworks, aiming to promote the safe and responsible development of LLM technology. This work seeks to inform researchers, practitioners, and policymakers, driving progress toward data security in LLMs.

Computer-using agents (CUAs), which can autonomously control computers to perform multi-step actions, might pose significant safety risks if misused. However, existing benchmarks mainly evaluate LMs in chatbots or simple tool use. To more comprehensively evaluate CUAs' misuse risks, we introduce a new benchmark: CUAHarm. CUAHarm consists of 104 expert-written realistic misuse risks, such as disabling firewalls, leaking data, or installing backdoors. We provide a sandbox with rule-based verifiable rewards to measure CUAs' success rates in executing these tasks (e.g., whether the firewall is indeed disabled), beyond refusal rates. We evaluate frontier LMs including GPT-5, Claude 4 Sonnet, Gemini 2.5 Pro, Llama-3.3-70B, and Mistral Large 2. Even without jailbreaking prompts, these frontier LMs comply with executing these malicious tasks at a high success rate (e.g., 90% for Gemini 2.5 Pro). Furthermore, while newer models are safer in previous safety benchmarks, their misuse risks as CUAs become even higher, e.g., Gemini 2.5 Pro is riskier than Gemini 1.5 Pro. Additionally, while these LMs are robust to common malicious prompts (e.g., creating a bomb) when acting as chatbots, they could still act unsafely as CUAs. We further evaluate a leading agentic framework (UI-TARS-1.5) and find that while it improves performance, it also amplifies misuse risks. To mitigate the misuse risks of CUAs, we explore using LMs to monitor CUAs' actions. We find monitoring unsafe computer-using actions is significantly harder than monitoring conventional unsafe chatbot responses. While monitoring chain-of-thoughts leads to modest gains, the average monitoring accuracy is only 77%. A hierarchical summarization strategy improves performance by up to 13%, a promising direction though monitoring remains unreliable. CUAHarm is released at https://github.com/db-ol/CUAHarm to facilitate further research.

Large language models (LLMs) are increasingly used to answer health-related questions and support healthcare workflows, yet evidence for their safety still relies heavily on static benchmarks that can rapidly become obsolete or be optimized against. Here we introduce a Dynamic, Automatic, and Systematic (DAS) red-teaming audit framework that continuously stress-tests LLMs for health across four safety-critical axes: robustness, privacy, bias/fairness, and hallucination/factual inaccuracies. Validated against board-certified clinicians with high concordance, a suite of adversarial agents autonomously mutates health-related test cases to uncover vulnerabilities in real time. Applying DAS to 15 proprietary and open-source LLMs revealed a profound gap between high static benchmark performance and low dynamic reliability--the "Benchmarking Gap". Despite median MedQA accuracy exceeding 80%, 94% of previously correct answers failed under dynamic robustness testing. This brittleness generalized to the realistic, open-ended HealthBench dataset, where top-tier models exhibited failure rates exceeding 70% and sharp shifts in model rankings across evaluations, suggesting that high scores on established static benchmarks may reflect superficial memorization. We observed similarly high failure rates across other domains: privacy leaks were elicited in 86% of scenarios, cognitive-bias priming altered recommendations in 81% of fairness tests, and hallucination rates exceeded 74% in widely used models. By converting LLM safety evaluation for health from a static checklist into a living adversarial audit, DAS provides a scalable framework for surfacing latent risks before such systems are deployed in consumer-facing health assistants, clinician-facing tools, and broader healthcare workflows. Code is available at https://github.com/JZPeterPan/DAS-Medical-Red-Teaming-Agents.

LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model's judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without reversing the model's internal belief of harmfulness. We also find that adversarially finetuning models to accept harmful instructions has minimal impact on the model's internal belief of harmfulness. These insights lead to a practical safety application: The model's latent harmfulness representation can serve as an intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs' internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety.

The proliferation of Low-Rank Adaptation (LoRA) models has democratized personalized text-to-image generation, enabling users to share lightweight models (e.g., personal portraits) on platforms like Civitai and Liblib. However, this "share-and-play" ecosystem introduces critical risks: benign LoRAs can be weaponized by adversaries to generate harmful content (e.g., political, defamatory imagery), undermining creator rights and platform safety. Existing defenses like concept-erasure methods focus on full diffusion models (DMs), neglecting LoRA's unique role as a modular adapter and its vulnerability to adversarial prompt engineering. To bridge this gap, we propose LoRAShield, the first data-free editing framework for securing LoRA models against misuse. Our platform-driven approach dynamically edits and realigns LoRA's weight subspace via adversarial optimization and semantic augmentation. Experimental results demonstrate that LoRAShield achieves remarkable effectiveness, efficiency, and robustness in blocking malicious generations without sacrificing the functionality of the benign task. By shifting the defense to platforms, LoRAShield enables secure, scalable sharing of personalized models, a critical step toward trustworthy generative ecosystems.

The increasing scale of AI workloads demands High-Performance Computing (HPC) infrastructure and training methodologies that are both scalable and sustainable. While Large Language Models (LLMs) demonstrate exceptional natural language capabilities, general-purpose models often lack the specialized domain knowledge necessary for effective cybersecurity analysis. We investigate Domain-Adaptive Continuous Pretraining (DAP) as a scalable, resource-efficient methodology for enhancing cybersecurity understanding in pretrained LLMs, implemented through a distributed Fully Sharded Data Parallel (FSDP) pipeline across multi-node GPU clusters. We systematically adapted three decoder-based architectures -- Llama-3.1-8B, DeepSeek-R1-Distill-Qwen-14B, and Llama-3.3-70B-Instruct -- using a curated 126-million-word cybersecurity corpus from standards, academic literature, and technical documentation. Evaluation across three cybersecurity benchmarks -- CTI-MCQ, CyberMetric, and SecEval -- demonstrates consistent improvements post-adaptation. Notably, our Llama-3.3-70B-Ins-DAP model achieves state-of-the-art performance with accuracies of 0.718, 0.933, and 0.864, respectively, surpassing parameter-efficient baselines and specialized models including Llama-Primus-Base (trained on 2.77 billion tokens) and Foundation-Sec-8B (trained on 5 billion tokens), despite utilizing only 118.8 million tokens -- representing a 23-to-42-fold reduction in training data. Targeted continuous pretraining via scalable HPC infrastructure enables effective cybersecurity domain adaptation with a substantially reduced computational and energy footprint, supporting specialized AI assistants in threat analysis, vulnerability assessment, and security documentation, while advancing sustainable and responsible AI development.

Large Language Models (LLMs) for code generation (i.e., Code LLMs) have demonstrated impressive capabilities in AI-assisted software development and testing. However, recent studies have shown that these models are prone to generating vulnerable or even malicious code under adversarial settings. Existing red-teaming approaches rely on extensive human effort, limiting their scalability and practicality, and generally overlook the interactive nature of real-world AI-assisted programming, which often unfolds over multiple turns. To bridge these gaps, we present RedCoder, a red-teaming agent that engages victim models in multi-turn conversation to elicit vulnerable code. The pipeline to construct RedCoder begins with a multi-agent gaming process that simulates adversarial interactions, yielding a set of prototype conversations and an arsenal of reusable attack strategies. We then fine-tune an LLM on these prototype conversations to serve as the backbone of RedCoder. Once deployed, RedCoder autonomously engages Code LLMs in multi-turn conversations, dynamically retrieving relevant strategies from the arsenal to steer the dialogue toward vulnerability-inducing outputs. Experiments across multiple Code LLMs show that our approach outperforms prior single-turn and multi-turn red-team methods in inducing vulnerabilities in code generation, offering a scalable and effective tool for evaluating the security boundaries of modern code-generation systems.