As LLM agents increasingly select tools autonomously, their choices among tools with different privileges become safety-relevant. However, prior tool-selection studies focus on safety-agnostic metadata preferences, leaving privilege-sensitive choices underexplored. To address this gap, we study over-privileged tool selection, in which an agent selects or escalates to a higher-privilege tool despite a sufficient lower-privilege alternative. We introduce ToolPrivBench to evaluate whether agents choose higher-privilege tools despite sufficient lower-privilege alternatives, measuring both initial selection and escalation after transient tool failures. Across eight domains and five recurring risk patterns, we find that over-privileged tool selection is common among mainstream LLM agents and is further amplified by transient failures. We further find that general safety alignment does not reliably transfer to least-privilege tool choice, while prompt-level controls provide only limited mitigation under transient failures. We therefore introduce a privilege-aware post-training defense that teaches agents to prefer sufficient lower-privilege tools and escalate only when necessary. Our mitigation experiments show that this defense substantially reduces unnecessary high-privilege tool use while preserving general capabilities.

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://anonymous.4open.science/r/FABLE-B59E.

Persistent vulnerabilities in safety alignment hinder the deployment of large language models (LLMs). Existing methods remain susceptible to jailbreak attacks, suggesting a fundamental, unaddressed flaw in current safety paradigms. In this work, we diagnose the root cause of this fragility, identifying a systemic issue we term Depth-wise Alignment Discrepancy. We find a fundamental misalignment: harmful vectors—latent representations predisposed to unsafe content—predominantly originate in the model’s lower layers, yet conventional alignment training concentrates its corrective gradients disproportionately on the top-most layers. This creates a brittle, "end-of-pipe" defense that is easily bypassed. To address this discrepancy, we propose SAGA, a framework that achieves robust safety through two synergistic innovations. First, it leverages high-entropy safety data, constructed via an iterative entropy ascent procedure, to provide the deep semantic signals necessary to reach the source of harmfulness. Second, it introduces a novel Synergistic Gradient Scaling (SGS) mechanism to explicitly reshape the gradient flow, ensuring these corrective signals are precisely applied to the identified vulnerable layers. Extensive experiments on five LLMs against six distinct jailbreak attacks demonstrate SAGA’s superiority, reducing attack success rates (ASR) by 21%–63% compared to state-of-the-art baselines. Our method preserves downstream task accuracy while introducing minimal computational overhead (<3%).

Despite substantial progress in safety alignment, aligned Large Language Models (LLMs) remain fragile under small parameter-space and activation-space perturbations, such as downstream fine-tuning and jailbreak attacks. To explain why local perturbations can trigger broad safety failures, we analyze safety-task neuron circuits from the perspectives of activation flow and circuit topology. We identify a critical topological fragility in safety alignment: In aligned models, safety-task neuron circuits exhibit a low-redundancy, isolated single-chain topology, in sharp contrast to the highly redundant dense-mesh topology of general-task circuits. This brittle topology creates rigid serial dependencies, whereby perturbing a small set of upstream neurons can disrupt downstream safety neurons and cascade into broad alignment failures. To mitigate this topological fragility, we propose Dense Safety Mesh (DSM), a topology-aware alignment framework that transforms brittle single-chain safety circuits into more redundant and resilient mesh-like structures. DSM employs structural adversarial masking to suppress dominant safety pathways, forcing safety circuits to develop diverse parallel pathways. Extensive experiments across seven models and multiple evaluations show that DSM improves robustness against perturbations attacks while maintaining general capabilities.

Prevailing safety alignment methods still leave Large Language Models (LLMs) vulnerable to sophisticated jailbreak attacks. To bolster defenses, explicit reasoning mechanisms like Safety-oriented Chain-of-Thought (SCoT) have emerged, significantly enhancing robustness. However, this transparency introduces a critical trade-off: the exposed reasoning process itself becomes a new attack surface, risking the leakage of harmful information and revealing the model’s safety logic to adversaries. This paper directly confronts this dilemma, asking: Can we achieve the full benefits of deliberative safety without the costs of explicit reasoning generation? We propose Safety Reasoning Internalization to make the deliberative process in SCoT “available but not visible”. This approach is grounded in a key theoretical insight: the corrective influence of an SCoT can be effectively approximated by a targeted, low-rank update to the model’s Feed-Forward Network (FFN) layers. We operationalize this through Hierarchical Internalization of Adversarially-Guided Reasoning (HIAR), a layer-wise safety alignment framework that internalizes safety reasoning into an implicit computational pathway using Low-Rank Adaptation (LoRA). HIAR enables the model to reach a safe conclusion within a single forward pass, entirely eliminating the need to generate vulnerable SCoT text. Extensive experiments on various LLMs demonstrate that HIAR achieves a 43% lower Attack Success Rate (ASR) against distinct jailbreak attacks compared to strong baselines.

To address the increasingly severe safety risk of large language models (LLMs), reasoning-based safety alignment methods have emerged. These methods overcome the limitations of ’shallow alignment’ by exposing the model’s Chain-of-Thought (CoT), enabling auditability of safety reasoning process through both training-phase supervision and post-generation verification. However, this transparency creates a critical vulnerability, a tension we define as the Security Auditability Dilemma: while explicit reasoning is a prerequisite for safety, its textual Auditable paradoxically transforms it into an optimization target for adaptive attackers and induces the model to unintentionally copy harmful content from its own reasoning context. To address this, we propose Auditable Latent CoT Alignment (ALCA), a framework that decouples internal reasoning from external output. ALCA shifts the safety deliberation process into a continuous latent space. This allows the safety reasoning process to guide the generation of harmless outputs, while eliminates the discrete textual surface that facilitates internal copying and adaptive attack. Yet, this process is not a black box. we introduce a restricted Self-Decoding mechanism that allows the model to reconstruct its latent reasoning into human-readable text for supervision under specific guidance. Extensive experiments show that ALCA achieves robustness alignment, reducing the success rate of adaptive jailbreak attacks by over 40% compared to strong baselines, while preserving performance. Our framework presents a path toward building LLMs that are both robustly secure and auditable.

The latent knowledge of large language models (LLMs) contains harmful or unethical content, which introduces significant security risks upon their widespread deployment. Conducting jailbreak attacks on LLMs can proactively identify vulnerabilities to enhance their security measures. However, previous jailbreak attacks primarily focus on single-turn dialogue scenarios, leaving vulnerabilities in multi-turn dialogue contexts inadequately explored. This paper investigates the resilience of black-box LLMs in multi-turn jailbreak attack scenarios from a novel interrogation perspective. We propose an optimal interrogation principle to conceal the jailbreak intent and introduce a multi-turn attack chain generation strategy called CoA. By employing two effective interrogation strategies tailored for LLMs, coupled with an interrogation history record management mechanis, it achieves a significant optimization of the attack process. Our approach enables the iterative generation of attack chains, offering a powerful tool for LLM red team testing. Experimental results demonstrate that LLMs exhibit insufficient resistance under multi-turn interrogation, with our method shows more advantages(ASR, 83% vs 64%). This work offers new insights into improving the safety of LLMs.

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, 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.