Large language models (LLMs) increasingly rely on explicit chain-of-thought (CoT) reasoning to solve complex tasks, yet the safety of the reasoning process itself remains largely unaddressed. Existing work on LLM safety focuses on content safety--detecting harmful, biased, or factually incorrect outputs -- and treats the reasoning chain as an opaque intermediate artifact. We identify reasoning safety as an orthogonal and equally critical security dimension: the requirement that a model's reasoning trajectory be logically consistent, computationally efficient, and resistant to adversarial manipulation. We make three contributions. First, we formally define reasoning safety and introduce a nine-category taxonomy of unsafe reasoning behaviors, covering input parsing errors, reasoning execution errors, and process management errors. Second, we conduct a large-scale prevalence study annotating 4111 reasoning chains from both natural reasoning benchmarks and four adversarial attack methods (reasoning hijacking and denial-of-service), confirming that all nine error types occur in practice and that each attack induces a mechanistically interpretable signature. Third, we propose a Reasoning Safety Monitor: an external LLM-based component that runs in parallel with the target model, inspects each reasoning step in real time via a taxonomy-embedded prompt, and dispatches an interrupt signal upon detecting unsafe behavior. Evaluation on a 450-chain static benchmark shows that our monitor achieves up to 84.88\% step-level localization accuracy and 85.37\% error-type classification accuracy, outperforming hallucination detectors and process reward model baselines by substantial margins. These results demonstrate that reasoning-level monitoring is both necessary and practically achievable, and establish reasoning safety as a foundational concern for the secure deployment of large reasoning models.

Large Language Models (LLMs) have become core cognitive components in modern artificial intelligence (AI) systems, combining internal knowledge with external context to perform complex tasks. However, LLMs typically treat all accessible data indiscriminately, lacking inherent awareness of knowledge ownership and access boundaries. This deficiency heightens risks of sensitive data leakage and adversarial manipulation, potentially enabling unauthorized system access and severe security crises. Existing protection strategies rely on rigid, uniform defense that prevent dynamic authorization. Structural isolation methods faces scalability bottlenecks, while prompt guidance methods struggle with fine-grained permissions distinctions. Here, we propose the Chain-of-Authorization (CoA) framework, a secure training and reasoning paradigm that internalizes authorization logic into LLMs' core capabilities. Unlike passive external defneses, CoA restructures the model's information flow: it embeds permission context at input and requires generating explicit authorization reasoning trajectory that includes resource review, identity resolution, and decision-making stages before final response. Through supervised fine-tuning on data covering various authorization status, CoA integrates policy execution with task responses, making authorization a causal prerequisite for substantive responses. Extensive evaluations show that CoA not only maintains comparable utility in authorized scenarios but also overcomes the cognitive confusion when permissions mismatches. It exhibits high rejection rates against various unauthorized and adversarial access. This mechanism leverages LLMs' reasoning capability to perform dynamic authorization, using natural language understanding as a proactive security mechanism for deploying reliable LLMs in modern AI systems.

Graph Neural Networks (GNNs) have achieved remarkable results in various tasks. Recent studies reveal that graph backdoor attacks can poison the GNN model to predict test nodes with triggers attached as the target class. However, apart from injecting triggers to training nodes, these graph backdoor attacks generally require altering the labels of trigger-attached training nodes into the target class, which is impractical in real-world scenarios. In this work, we focus on the clean-label graph backdoor attack, a realistic but understudied topic where training labels are not modifiable. According to our preliminary analysis, existing graph backdoor attacks generally fail under the clean-label setting. Our further analysis identifies that the core failure of existing methods lies in their inability to poison the prediction logic of GNN models, leading to the triggers being deemed unimportant for prediction. Therefore, we study a novel problem of effective clean-label graph backdoor attacks by poisoning the inner prediction logic of GNN models. We propose BA-Logic to solve the problem by coordinating a poisoned node selector and a logic-poisoning trigger generator. Extensive experiments on real-world datasets demonstrate that our method effectively enhances the attack success rate and surpasses state-of-the-art graph backdoor attack competitors under clean-label settings. Our code is available at https://anonymous.4open.science/r/BA-Logic

Video-LLMs are increasingly deployed in safety-critical applications but are vulnerable to Energy-Latency Attacks (ELAs) that exhaust computational resources. Current image-centric methods fail because temporal aggregation mechanisms dilute individual frame perturbations. Additionally, real-time demands make instance-wise optimization impractical for continuous video streams. We introduce VidDoS, which is the first universal ELA framework tailored for Video-LLMs. Our method leverages universal optimization to create instance-agnostic triggers that require no inference-time gradient calculation. We achieve this through $\textit{masked teacher forcing}$ to steer models toward expensive target sequences, combined with a $\textit{refusal penalty}$ and $\textit{early-termination suppression}$ to override conciseness priors. Testing across three mainstream Video-LLMs and three video datasets, which include video question answering and autonomous driving scenarios, shows extreme degradation. VidDoS induces a token expansion of more than 205$\times$ and inflates the inference latency by more than 15$\times$ relative to clean baselines. Simulations of real-time autonomous driving streams further reveal that this induced latency leads to critical safety violations. We urge the community to recognize and mitigate these high-hazard ELA in Video-LLMs.

Recent intelligent systems integrate powerful Large Language Models (LLMs) through APIs, but their trustworthiness may be critically undermined by targeted attacks like backdoor and prompt injection attacks, which secretly force LLMs to generate specific malicious sequences. Existing defensive approaches for such threats typically rely on high access rights, impose prohibitive costs, and hinder normal inference, rendering them impractical for real-world scenarios. To solve these limitations, we introduce DualSentinel, a lightweight and unified defense framework that can accurately and promptly detect the activation of targeted attacks alongside the LLM generation process. We first identify a characteristic of compromised LLMs, termed Entropy Lull: when a targeted attack successfully hijacks the generation process, the LLM exhibits a distinct period of abnormally low and stable token probability entropy, indicating it is following a fixed path rather than making creative choices. DualSentinel leverages this pattern by developing an innovative dual-check approach. It first employs a magnitude and trend-aware monitoring method to proactively and sensitively flag an entropy lull pattern at runtime. Upon such flagging, it triggers a lightweight yet powerful secondary verification based on task-flipping. An attack is confirmed only if the entropy lull pattern persists across both the original and the flipped task, proving that the LLM's output is coercively controlled. Extensive evaluations show that DualSentinel is both highly effective (superior detection accuracy with near-zero false positives) and remarkably efficient (negligible additional cost), offering a truly practical path toward securing deployed LLMs. The source code can be accessed at https://doi.org/10.5281/zenodo.18479273.

Fine-tuning large language models (LLMs) on custom datasets has become a standard approach for adapting these models to specific domains and applications. However, recent studies have shown that such fine-tuning can lead to significant degradation in the model's safety. Existing defense methods operate at the sample level and often suffer from an unsatisfactory trade-off between safety and utility. To address this limitation, we perform a systematic token-level diagnosis of safety degradation during fine-tuning. Based on this, we propose token-level data selection for safe LLM fine-tuning (TOSS), a novel framework that quantifies the safety risk of each token by measuring the loss difference between a safety-degraded model and a utility-oriented model. This token-level granularity enables accurate identification and removal of unsafe tokens, thereby preserving valuable task-specific information. In addition, we introduce a progressive refinement strategy, TOSS-Pro, which iteratively enhances the safety-degraded model's ability to identify unsafe tokens. Extensive experiments demonstrate that our approach robustly safeguards LLMs during fine-tuning while achieving superior downstream task performance, significantly outperforming existing sample-level defense methods. Our code is available at https://github.com/Polly-LYP/TOSS.

The evolution of Large Language Models (LLMs) into agentic systems that perform autonomous reasoning and tool use has created significant intellectual property (IP) value. We demonstrate that these systems are highly vulnerable to imitation attacks, where adversaries steal proprietary capabilities by training imitation models on victim outputs. Crucially, existing LLM watermarking techniques fail in this domain because real-world agentic systems often operate as grey boxes, concealing the internal reasoning traces required for verification. This paper presents AGENTWM, the first watermarking framework designed specifically for agentic models. AGENTWM exploits the semantic equivalence of action sequences, injecting watermarks by subtly biasing the distribution of functionally identical tool execution paths. This mechanism allows AGENTWM to embed verifiable signals directly into the visible action trajectory while remaining indistinguishable to users. We develop an automated pipeline to generate robust watermark schemes and a rigorous statistical hypothesis testing procedure for verification. Extensive evaluations across three complex domains demonstrate that AGENTWM achieves high detection accuracy with negligible impact on agent performance. Our results confirm that AGENTWM effectively protects agentic IP against adaptive adversaries, who cannot remove the watermarks without severely degrading the stolen model's utility.

Vision language models (VLMs) extend the reasoning capabilities of large language models (LLMs) to cross-modal settings, yet remain highly vulnerable to multimodal jailbreak attacks. Existing defenses predominantly rely on safety fine-tuning or aggressive token manipulations, incurring substantial training costs or significantly degrading utility. Recent research shows that LLMs inherently recognize unsafe content in text, and the incorporation of visual inputs in VLMs frequently dilutes risk-related signals. Motivated by this, we propose Risk Awareness Injection (RAI), a lightweight and training-free framework for safety calibration that restores LLM-like risk recognition by amplifying unsafe signals in VLMs. Specifically, RAI constructs an Unsafe Prototype Subspace from language embeddings and performs targeted modulation on selected high-risk visual tokens, explicitly activating safety-critical signals within the cross-modal feature space. This modulation restores the model's LLM-like ability to detect unsafe content from visual inputs, while preserving the semantic integrity of original tokens for cross-modal reasoning. Extensive experiments across multiple jailbreak and utility benchmarks demonstrate that RAI substantially reduces attack success rate without compromising task performance.

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.

The agent-tool communication loop is a critical attack surface in modern Large Language Model (LLM) agents. Existing Denial-of-Service (DoS) attacks, primarily triggered via user prompts or injected retrieval-augmented generation (RAG) context, are ineffective for this new paradigm. They are fundamentally single-turn and often lack a task-oriented approach, making them conspicuous in goal-oriented workflows and unable to exploit the compounding costs of multi-turn agent-tool interactions. We introduce a stealthy, multi-turn economic DoS attack that operates at the tool layer under the guise of a correctly completed task. Our method adjusts text-visible fields and a template-governed return policy in a benign, Model Context Protocol (MCP)-compatible tool server, optimizing these edits with a Monte Carlo Tree Search (MCTS) optimizer. These adjustments leave function signatures unchanged and preserve the final payload, steering the agent into prolonged, verbose tool-calling sequences using text-only notices. This compounds costs across turns, escaping single-turn caps while keeping the final answer correct to evade validation. Across six LLMs on the ToolBench and BFCL benchmarks, our attack expands tasks into trajectories exceeding 60,000 tokens, inflates costs by up to 658x, and raises energy by 100-560x. It drives GPU KV cache occupancy from <1% to 35-74% and cuts co-running throughput by approximately 50%. Because the server remains protocol-compatible and task outcomes are correct, conventional checks fail. These results elevate the agent-tool interface to a first-class security frontier, demanding a paradigm shift from validating final answers to monitoring the economic and computational cost of the entire agentic process.

Watermarking has emerged as a pivotal solution for content traceability and intellectual property protection in Large Vision-Language Models (LVLMs). However, vision-agnostic watermarks introduce visually irrelevant tokens and disrupt visual grounding by enforcing indiscriminate pseudo-random biases, while some semantic-aware methods incur prohibitive inference latency due to rejection sampling. In this paper, we propose the VIsual Semantic Adaptive Watermark (VISA-Mark), a novel framework that embeds detectable signals while strictly preserving visual fidelity. Our approach employs a lightweight, efficiently trained prefix-tuner to extract dynamic Visual-Evidence Weights, which quantify the evidentiary support for candidate tokens based on the visual input. These weights guide an adaptive vocabulary partitioning and logits perturbation mechanism, concentrating watermark strength specifically on visually-supported tokens. By actively aligning the watermark with visual evidence, VISA-Mark effectively maintains visual fidelity. Empirical results confirm that VISA-Mark outperforms conventional methods with a 7.8% improvement in visual consistency (Chair-I) and superior semantic fidelity. The framework maintains highly competitive detection accuracy (96.88% AUC) and robust attack resilience (99.3%) without sacrificing inference efficiency, effectively establishing a new standard for reliability-preserving multimodal watermarking.

Reasoning Large Language Models (RLLMs) excelling in complex tasks present unique challenges for digital watermarking, as existing methods often disrupt logical coherence or incur high computational costs. Token-based watermarking techniques can corrupt the reasoning flow by applying pseudo-random biases, while semantic-aware approaches improve quality but introduce significant latency or require auxiliary models. This paper introduces ReasonMark, a novel watermarking framework specifically designed for reasoning-intensive LLMs. Our approach decouples generation into an undisturbed Thinking Phase and a watermarked Answering Phase. We propose a Criticality Score to identify semantically pivotal tokens from the reasoning trace, which are distilled into a Principal Semantic Vector (PSV). The PSV then guides a semantically-adaptive mechanism that modulates watermark strength based on token-PSV alignment, ensuring robustness without compromising logical integrity. Extensive experiments show ReasonMark surpasses state-of-the-art methods by reducing text Perplexity by 0.35, increasing translation BLEU score by 0.164, and raising mathematical accuracy by 0.67 points. These advancements are achieved alongside a 0.34% higher watermark detection AUC and stronger robustness to attacks, all with a negligible increase in latency. This work enables the traceable and trustworthy deployment of reasoning LLMs in real-world applications.

Deep learning advances have enabled accurate six-degree-of-freedom (6DoF) object pose estimation, widely used in robotics, AR/VR, and autonomous systems. However, backdoor attacks pose significant security risks. While most research focuses on 2D vision, 6DoF pose estimation remains largely unexplored. Unlike traditional backdoors that only change classes, 6DoF attacks must control continuous parameters like translation and rotation, rendering 2D methods inapplicable. We propose 6DAttack, a framework using 3D object triggers to induce controlled erroneous poses while maintaining normal behavior. Evaluations on PVNet, DenseFusion, and PoseDiffusion across LINEMOD, YCB-Video, and CO3D show high attack success rates (ASRs) without compromising clean performance. Backdoored models achieve up to 100% clean ADD accuracy and 100% ASR, with triggered samples reaching 97.70% ADD-P. Furthermore, a representative defense remains ineffective. Our findings reveal a serious, underexplored threat to 6DoF pose estimation.

Large Language Models (LLMs) are increasingly used in agentic systems, where their interactions with diverse tools and environments create complex, multi-stage safety challenges. However, existing benchmarks mostly rely on static, single-turn assessments that miss vulnerabilities from adaptive, long-chain attacks. To fill this gap, we introduce DREAM, a framework for systematic evaluation of LLM agents against dynamic, multi-stage attacks. At its core, DREAM uses a Cross-Environment Adversarial Knowledge Graph (CE-AKG) to maintain stateful, cross-domain understanding of vulnerabilities. This graph guides a Contextualized Guided Policy Search (C-GPS) algorithm that dynamically constructs attack chains from a knowledge base of 1,986 atomic actions across 349 distinct digital environments. Our evaluation of 12 leading LLM agents reveals a critical vulnerability: these attack chains succeed in over 70% of cases for most models, showing the power of stateful, cross-environment exploits. Through analysis of these failures, we identify two key weaknesses in current agents: contextual fragility, where safety behaviors fail to transfer across environments, and an inability to track long-term malicious intent. Our findings also show that traditional safety measures, such as initial defense prompts, are largely ineffective against attacks that build context over multiple interactions. To advance agent safety research, we release DREAM as a tool for evaluating vulnerabilities and developing more robust defenses.

It has been demonstrated that adversarial graphs, i.e., graphs with imperceptible perturbations, can cause deep graph models to fail on classification tasks. In this work, we extend the concept of adversarial graphs to the community detection problem, which is more challenging. We propose novel attack and defense techniques for community detection problem, with the objective of hiding targeted individuals from detection models and enhancing the robustness of community detection models, respectively. These techniques have many applications in real-world scenarios, for example, protecting personal privacy in social networks and understanding camouflage patterns in transaction networks. To simulate interactive attack and defense behaviors, we further propose a game-theoretic framework, called CD-GAME. One player is a graph attacker, while the other player is a Rayleigh Quotient defender. The CD-GAME models the mutual influence and feedback mechanisms between the attacker and the defender, revealing the dynamic evolutionary process of the game. Both players dynamically update their strategies until they reach the Nash equilibrium. Extensive experiments demonstrate the effectiveness of our proposed attack and defense methods, and both outperform existing baselines by a significant margin. Furthermore, CD-GAME provides valuable insights for understanding interactive attack and defense scenarios in community detection problems. We found that in traditional single-step attack or defense, attacker tends to employ strategies that are most effective, but are easily detected and countered by defender. When the interactive game reaches a Nash equilibrium, attacker adopts more imperceptible strategies that can still achieve satisfactory attack effectiveness even after defense.

As LLMs become more common, non-expert users can pose risks, prompting extensive research into jailbreak attacks. However, most existing black-box jailbreak attacks rely on hand-crafted heuristics or narrow search spaces, which limit scalability. Compared with prior attacks, we propose Game-Theory Attack (GTA), an scalable black-box jailbreak framework. Concretely, we formalize the attacker's interaction against safety-aligned LLMs as a finite-horizon, early-stoppable sequential stochastic game, and reparameterize the LLM's randomized outputs via quantal response. Building on this, we introduce a behavioral conjecture "template-over-safety flip": by reshaping the LLM's effective objective through game-theoretic scenarios, the originally safety preference may become maximizing scenario payoffs within the template, which weakens safety constraints in specific contexts. We validate this mechanism with classical game such as the disclosure variant of the Prisoner's Dilemma, and we further introduce an Attacker Agent that adaptively escalates pressure to increase the ASR. Experiments across multiple protocols and datasets show that GTA achieves over 95% ASR on LLMs such as Deepseek-R1, while maintaining efficiency. Ablations over components, decoding, multilingual settings, and the Agent's core model confirm effectiveness and generalization. Moreover, scenario scaling studies further establish scalability. GTA also attains high ASR on other game-theoretic scenarios, and one-shot LLM-generated variants that keep the model mechanism fixed while varying background achieve comparable ASR. Paired with a Harmful-Words Detection Agent that performs word-level insertions, GTA maintains high ASR while lowering detection under prompt-guard models. Beyond benchmarks, GTA jailbreaks real-world LLM applications and reports a longitudinal safety monitoring of popular HuggingFace LLMs.

Large Language Model (LLM)-based agents with function-calling capabilities are increasingly deployed, but remain vulnerable to Indirect Prompt Injection (IPI) attacks that hijack their tool calls. In response, numerous IPI-centric defense frameworks have emerged. However, these defenses are fragmented, lacking a unified taxonomy and comprehensive evaluation. In this Systematization of Knowledge (SoK), we present the first comprehensive analysis of IPI-centric defense frameworks. We introduce a comprehensive taxonomy of these defenses, classifying them along five dimensions. We then thoroughly assess the security and usability of representative defense frameworks. Through analysis of defensive failures in the assessment, we identify six root causes of defense circumvention. Based on these findings, we design three novel adaptive attacks that significantly improve attack success rates targeting specific frameworks, demonstrating the severity of the flaws in these defenses. Our paper provides a foundation and critical insights for the future development of more secure and usable IPI-centric agent defense frameworks.

Membership inference attacks (MIAs) on large language models (LLMs) pose significant privacy risks across various stages of model training. Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have brought a profound paradigm shift in LLM training, particularly for complex reasoning tasks. However, the on-policy nature of RLVR introduces a unique privacy leakage pattern: since training relies on self-generated responses without fixed ground-truth outputs, membership inference must now determine whether a given prompt (independent of any specific response) is used during fine-tuning. This creates a threat where leakage arises not from answer memorization. To audit this novel privacy risk, we propose Divergence-in-Behavior Attack (DIBA), the first membership inference framework specifically designed for RLVR. DIBA shifts the focus from memorization to behavioral change, leveraging measurable shifts in model behavior across two axes: advantage-side improvement (e.g., correctness gain) and logit-side divergence (e.g., policy drift). Through comprehensive evaluations, we demonstrate that DIBA significantly outperforms existing baselines, achieving around 0.8 AUC and an order-of-magnitude higher TPR@0.1%FPR. We validate DIBA's superiority across multiple settings--including in-distribution, cross-dataset, cross-algorithm, black-box scenarios, and extensions to vision-language models. Furthermore, our attack remains robust under moderate defensive measures. To the best of our knowledge, this is the first work to systematically analyze privacy vulnerabilities in RLVR, revealing that even in the absence of explicit supervision, training data exposure can be reliably inferred through behavioral traces.

Large language models (LLMs) have achieved remarkable success across a wide range of natural language processing tasks, demonstrating human-level performance in text generation, reasoning, and question answering. However, training such models requires substantial computational resources, large curated datasets, and sophisticated alignment procedures. As a result, they constitute highly valuable intellectual property (IP) assets that warrant robust protection mechanisms. Existing IP protection approaches suffer from critical limitations. Model fingerprinting techniques can identify model architectures but fail to establish ownership of specific model instances. In contrast, traditional backdoor-based watermarking methods embed behavioral anomalies that can be easily removed through common post-processing operations such as fine-tuning or knowledge distillation. We propose SEAL, a subspace-anchored watermarking framework that embeds multi-bit signatures directly into the model's latent representational space, supporting both white-box and black-box verification scenarios. Our approach leverages model editing techniques to align the hidden representations of selected anchor samples with predefined orthogonal bit vectors. This alignment embeds the watermark while preserving the model's original factual predictions, rendering the watermark functionally harmless and stealthy. We conduct comprehensive experiments on multiple benchmark datasets and six prominent LLMs, comparing SEAL with 11 existing fingerprinting and watermarking methods to demonstrate its superior effectiveness, fidelity, efficiency, and robustness. Furthermore, we evaluate SEAL under potential knowledgeable attacks and show that it maintains strong verification performance even when adversaries possess knowledge of the watermarking mechanism and the embedded signatures.

The increasing realism of generated images has raised significant concerns about their potential misuse, necessitating robust detection methods. Current approaches mainly rely on training binary classifiers, which depend heavily on the quantity and quality of available generated images. In this work, we propose a novel framework that exploits geometric differences between the data manifolds of natural and generated images. To exploit this difference, we employ a pair of functions engineered to yield consistent outputs for natural images but divergent outputs for generated ones, leveraging the property that their gradients reside in mutually orthogonal subspaces. This design enables a simple yet effective detection method: an image is identified as generated if a transformation along its data manifold induces a significant change in the loss value of a self-supervised model pre-trained on natural images. Further more, to address diminishing manifold disparities in advanced generative models, we leverage normalizing flows to amplify detectable differences by extruding generated images away from the natural image manifold. Extensive experiments demonstrate the efficacy of this method. Code is available at https://github.com/tmlr-group/ConV.