Structured Security Auditing and Robustness Enhancement for Untrusted Agent Skills

Agent Skills package SKILL.md files, scripts, reference documents, and repository context into reusable capability units, turning pre-load auditing from single-prompt filtering into cross-file security review. Existing guardrails often flag risk but recover malicious intent inconsistently under semantics-preserving rewrites. This paper formulates pre-load auditing for untrusted Agent Skills as a robust three-way classification task and introduces SkillGuard-Robust, which combines role-aware evidence extraction, selective semantic verification, and consistency-preserving adjudication. We evaluate SkillGuard-Robust on SkillGuardBench and two public-ecosystem extensions through five large evaluation views ranging from 254 to 404 packages. On the 404-package held-out aggregate, SkillGuard-Robust reaches 97.30% overall exact match, 98.33% malicious-risk recall, and 98.89% attack exact consistency. On the 254-package external-ecosystem view, it reaches 99.66%, 100.00%, and 100.00%, respectively. These results support a bounded conclusion: factorized package auditing materially improves frozen and public-ecosystem robustness, while harsher external-source transfer remains an open challenge.

RouteGuard: Internal-Signal Detection of Skill Poisoning in LLM Agents

Agent skills introduce a new and more severe form of indirect injection for LLM agents: unlike traditional indirect prompt injection, attackers can hide malicious instructions inside a dense, action-oriented skill that already functions as a legitimate instruction source. We study pre-execution skill-poison detection and show that successful skill poisoning induces a structured internal effect, attention hijacking, in which response-time attention shifts from trusted context to malicious skill spans and drives harmful behavior. Motivated by this mechanism, we propose RouteGuard, a frozen-backbone detector that combines response-conditioned attention and hidden-state alignment through reliability-gated late fusion. Across both real and synthetic open-source skill benchmarks, RouteGuard is consistently the strongest or most robust detector; on the critical Skill-Inject channel slice, it reaches 0.8834 F1 and recovers 90.51% of description attacks missed by lexical screening, showing that defending against skill poisoning requires internal-signal detection rather than text-only filtering

FABLE: Fine-grained Fact Anchoring for Unstructured Model Editing

Unstructured model editing aims to update models with real-world text, yet existing methods often memorize text holistically without reliable fine-grained fact access. To address this, we propose FABLE, a hierarchical framework that decouples fine-grained fact injection from holistic text generation. FABLE follows a two-stage, fact-first strategy: discrete facts are anchored in shallow layers, followed by minimal updates to deeper layers to produce coherent text. This decoupling resolves the mismatch between holistic recall and fine-grained fact access, reflecting the unidirectional Transformer flow in which surface-form generation amplifies rather than corrects underlying fact representations. We also introduce UnFine, a diagnostic benchmark with fine-grained question-answer pairs and fact-level metrics for systematic evaluation. Experiments show that FABLE substantially improves fine-grained question answering while maintaining state-of-the-art holistic editing performance. Our code is publicly available at https://github.com/caskcsg/FABLE.

HearSay Benchmark: Do Audio LLMs Leak What They Hear?

While Audio Large Language Models (ALLMs) have achieved remarkable progress in understanding and generation, their potential privacy implications remain largely unexplored. This paper takes the first step to investigate whether ALLMs inadvertently leak user privacy solely through acoustic voiceprints and introduces $\textit{HearSay}$, a comprehensive benchmark constructed from over 22,000 real-world audio clips. To ensure data quality, the benchmark is meticulously curated through a rigorous pipeline involving automated profiling and human verification, guaranteeing that all privacy labels are grounded in factual records. Extensive experiments on $\textit{HearSay}$ yield three critical findings: $\textbf{Significant Privacy Leakage}$: ALLMs inherently extract private attributes from voiceprints, reaching 92.89% accuracy on gender and effectively profiling social attributes. $\textbf{Insufficient Safety Mechanisms}$: Alarmingly, existing safeguards are severely inadequate; most models fail to refuse privacy-intruding requests, exhibiting near-zero refusal rates for physiological traits. $\textbf{Reasoning Amplifies Risk}$: Chain-of-Thought (CoT) reasoning exacerbates privacy risks in capable models by uncovering deeper acoustic correlations. These findings expose critical vulnerabilities in ALLMs, underscoring the urgent need for targeted privacy alignment. The codes and dataset are available at https://github.com/JinWang79/HearSay_Benchmark

Exploiting Synergistic Cognitive Biases to Bypass Safety in LLMs

Large Language Models (LLMs) demonstrate impressive capabilities across a wide range of tasks, yet their safety mechanisms remain susceptible to adversarial attacks that exploit cognitive biases -- systematic deviations from rational judgment. Unlike prior jailbreaking approaches focused on prompt engineering or algorithmic manipulation, this work highlights the overlooked power of multi-bias interactions in undermining LLM safeguards. We propose CognitiveAttack, a novel red-teaming framework that systematically leverages both individual and combined cognitive biases. By integrating supervised fine-tuning and reinforcement learning, CognitiveAttack generates prompts that embed optimized bias combinations, effectively bypassing safety protocols while maintaining high attack success rates. Experimental results reveal significant vulnerabilities across 30 diverse LLMs, particularly in open-source models. CognitiveAttack achieves a substantially higher attack success rate compared to the SOTA black-box method PAP (60.1% vs. 31.6%), exposing critical limitations in current defense mechanisms. These findings highlight multi-bias interactions as a powerful yet underexplored attack vector. This work introduces a novel interdisciplinary perspective by bridging cognitive science and LLM safety, paving the way for more robust and human-aligned AI systems.

Paper Summary Attack: Jailbreaking LLMs through LLM Safety Papers

The safety of large language models (LLMs) has garnered significant research attention. In this paper, we argue that previous empirical studies demonstrate LLMs exhibit a propensity to trust information from authoritative sources, such as academic papers, implying new possible vulnerabilities. To verify this possibility, a preliminary analysis is designed to illustrate our two findings. Based on this insight, a novel jailbreaking method, Paper Summary Attack (\llmname{PSA}), is proposed. It systematically synthesizes content from either attack-focused or defense-focused LLM safety paper to construct an adversarial prompt template, while strategically infilling harmful query as adversarial payloads within predefined subsections. Extensive experiments show significant vulnerabilities not only in base LLMs, but also in state-of-the-art reasoning model like Deepseek-R1. PSA achieves a 97\% attack success rate (ASR) on well-aligned models like Claude3.5-Sonnet and an even higher 98\% ASR on Deepseek-R1. More intriguingly, our work has further revealed diametrically opposed vulnerability bias across different base models, and even between different versions of the same model, when exposed to either attack-focused or defense-focused papers. This phenomenon potentially indicates future research clues for both adversarial methodologies and safety alignment.Code is available at https://github.com/233liang/Paper-Summary-Attack

LyapLock: Bounded Knowledge Preservation in Sequential Large Language Model Editing

Large Language Models often contain factually incorrect or outdated knowledge, giving rise to model editing methods for precise knowledge updates. However, current mainstream locate-then-edit approaches exhibit a progressive performance decline during sequential editing, due to inadequate mechanisms for long-term knowledge preservation. To tackle this, we model the sequential editing as a constrained stochastic programming. Given the challenges posed by the cumulative preservation error constraint and the gradually revealed editing tasks, \textbf{LyapLock} is proposed. It integrates queuing theory and Lyapunov optimization to decompose the long-term constrained programming into tractable stepwise subproblems for efficient solving. This is the first model editing framework with rigorous theoretical guarantees, achieving asymptotic optimal editing performance while meeting the constraints of long-term knowledge preservation. Experimental results show that our framework scales sequential editing capacity to over 10,000 edits while stabilizing general capabilities and boosting average editing efficacy by 11.89\% over SOTA baselines. Furthermore, it can be leveraged to enhance the performance of baseline methods. Our code is released on https://github.com/caskcsg/LyapLock.

The Dark Side of Trust: Authority Citation-Driven Jailbreak Attacks on Large Language Models

The widespread deployment of large language models (LLMs) across various domains has showcased their immense potential while exposing significant safety vulnerabilities. A major concern is ensuring that LLM-generated content aligns with human values. Existing jailbreak techniques reveal how this alignment can be compromised through specific prompts or adversarial suffixes. In this study, we introduce a new threat: LLMs' bias toward authority. While this inherent bias can improve the quality of outputs generated by LLMs, it also introduces a potential vulnerability, increasing the risk of producing harmful content. Notably, the biases in LLMs is the varying levels of trust given to different types of authoritative information in harmful queries. For example, malware development often favors trust GitHub. To better reveal the risks with LLM, we propose DarkCite, an adaptive authority citation matcher and generator designed for a black-box setting. DarkCite matches optimal citation types to specific risk types and generates authoritative citations relevant to harmful instructions, enabling more effective jailbreak attacks on aligned LLMs.Our experiments show that DarkCite achieves a higher attack success rate (e.g., LLama-2 at 76% versus 68%) than previous methods. To counter this risk, we propose an authenticity and harm verification defense strategy, raising the average defense pass rate (DPR) from 11% to 74%. More importantly, the ability to link citations to the content they encompass has become a foundational function in LLMs, amplifying the influence of LLMs' bias toward authority.

AdaPPA: Adaptive Position Pre-Fill Jailbreak Attack Approach Targeting LLMs

Jailbreak vulnerabilities in Large Language Models (LLMs) refer to methods that extract malicious content from the model by carefully crafting prompts or suffixes, which has garnered significant attention from the research community. However, traditional attack methods, which primarily focus on the semantic level, are easily detected by the model. These methods overlook the difference in the model's alignment protection capabilities at different output stages. To address this issue, we propose an adaptive position pre-fill jailbreak attack approach for executing jailbreak attacks on LLMs. Our method leverages the model's instruction-following capabilities to first output pre-filled safe content, then exploits its narrative-shifting abilities to generate harmful content. Extensive black-box experiments demonstrate our method can improve the attack success rate by 47% on the widely recognized secure model (Llama2) compared to existing approaches. Our code can be found at: https://github.com/Yummy416/AdaPPA.

Enhancing Cross-Prompt Transferability in Vision-Language Models through Contextual Injection of Target Tokens

Vision-language models (VLMs) seamlessly integrate visual and textual data to perform tasks such as image classification, caption generation, and visual question answering. However, adversarial images often struggle to deceive all prompts effectively in the context of cross-prompt migration attacks, as the probability distribution of the tokens in these images tends to favor the semantics of the original image rather than the target tokens. To address this challenge, we propose a Contextual-Injection Attack (CIA) that employs gradient-based perturbation to inject target tokens into both visual and textual contexts, thereby improving the probability distribution of the target tokens. By shifting the contextual semantics towards the target tokens instead of the original image semantics, CIA enhances the cross-prompt transferability of adversarial images.Extensive experiments on the BLIP2, InstructBLIP, and LLaVA models show that CIA outperforms existing methods in cross-prompt transferability, demonstrating its potential for more effective adversarial strategies in VLMs.