Existing research on LLM agent security mainly focuses on prompt injection and unsafe input/output behaviors. However, as agents increasingly rely on third-party tools and MCP servers, a new class of supply-chain threats has emerged, where malicious behaviors are embedded in seemingly benign tools, silently hijacking agent execution, leaking sensitive data, or triggering unauthorized actions. Despite their growing impact, there is currently no comprehensive benchmark for evaluating such threats. To bridge this gap, we introduce SC-Inject-Bench, a large-scale benchmark comprising over 10,000 malicious MCP tools grounded in a taxonomy of 25+ attack types derived from MITRE ATT&CK targeting supply-chain threats. We observe that existing MCP scanners and semantic guardrails perform poorly on this benchmark. Motivated by this finding, we propose ShieldNet, a network-level guardrail framework that detects supply-chain poisoning by observing real network interactions rather than surface-level tool traces. ShieldNet integrates a man-in-the-middle (MITM) proxy and an event extractor to identify critical network behaviors, which are then processed by a lightweight classifier for attack detection. Extensive experiments show that ShieldNet achieves strong detection performance (up to 0.995 F-1 with only 0.8% false positives) while introducing little runtime overhead, substantially outperforming existing MCP scanners and LLM-based guardrails.

The rapid deployment of open-source frameworks has significantly advanced the development of modern multi-agent systems. However, expanded action spaces, including uncontrolled privilege exposure and hidden inter-system interactions, pose severe security challenges. Specifically, Indirect Prompt Injections (IPI), which conceal malicious instructions within third-party content, can trigger unauthorized actions such as data exfiltration during normal operations. While current security evaluations predominantly rely on isolated single-turn benchmarks, the systemic vulnerabilities of these agents within complex dynamic environments remain critically underexplored. To bridge this gap, we systematically evaluate six defense strategies against four sophisticated IPI attack vectors across nine LLM backbones. Crucially, we conduct our evaluation entirely within dynamic multi-step tool-calling environments to capture the true attack surface of modern autonomous agents. Moving beyond binary success rates, our multidimensional analysis reveals a pronounced fragility. Advanced injections successfully bypass nearly all baseline defenses, and some surface-level mitigations even produce counterproductive side effects. Furthermore, while agents execute malicious instructions almost instantaneously, their internal states exhibit abnormally high decision entropy. Motivated by this latent hesitation, we investigate Representation Engineering (RepE) as a robust detection strategy. By extracting hidden states at the tool-input position, we revealed that the RepE-based circuit breaker successfully identifies and intercepts unauthorized actions before the agent commits to them, achieving high detection accuracy across diverse LLM backbones. This study exposes the limitations of current IPI defenses and provides a highly practical paradigm for building resilient multi-agent architectures.

Reasoning language models (RLMs) are increasingly used in programming. Yet, even state-of-the-art RLMs frequently introduce critical security vulnerabilities in generated code. Prior training-based approaches for secure code generation face a critical limitation that prevents their direct application to RLMs: they rely on costly, manually curated security datasets covering only a limited set of vulnerabilities. At the inference level, generic security reminders consistently degrade functional correctness while triggering only shallow ad-hoc vulnerability analysis. To address these problems, we present SecPI, a fine-tuning pipeline that teaches RLMs to internalize structured security reasoning, producing secure code by default without any security instructions at inference time. SecPI filters existing general-purpose coding datasets for security-relevant tasks using an LLM-based classifier, generates high-quality security reasoning traces with a teacher model guided by a structured prompt that systematically enumerates relevant CWEs and mitigations, and fine-tunes the target model on pairs of inputs with no security prompt and teacher reasoning traces -- as a result, the model learns to reason about security autonomously rather than in response to explicit instructions. An extensive evaluation on security benchmarks with state-of-the-art open-weight reasoning models validates the effectiveness of our approach. For instance, SecPI improves the percentage of functionally correct and secure generations for QwQ 32B from 48.2% to 62.2% (+14.0 points) on CWEval and from 18.2% to 22.0% on BaxBench. Further investigation also reveals strong cross-CWE and cross-language generalization beyond training vulnerabilities. Even when trained only on injection-related CWEs, QwQ 32B generates correct and secure code 9.9% more frequently on held-out memory-safety CWEs.

Recent Tracking-by-Query-Propagation (TBP) methods have advanced Multi-Object Tracking (MOT) by enabling end-to-end (E2E) pipelines with long-range temporal modeling. However, this reliance on query propagation introduces unexplored architectural vulnerabilities to adversarial attacks. We present FADE, a novel attack framework designed to exploit these specific vulnerabilities. FADE employs two attack strategies targeting core TBP mechanisms: (i) Temporal Query Flooding: Generates spurious temporally consistent track queries to exhaust the tracker's limited query budget, forcing it to terminate valid tracks. (ii) Temporal Memory Corruption: Directly attacks the query updater's memory by severing temporal links via state de-correlation and erasing the learned feature identity of matched tracks. Furthermore, we introduce a differentiable pipeline to optimize these attacks for physical-world realizability by leveraging simulations of advanced perception sensor spoofing. Experiments on MOT17 and MOT20 benchmarks demonstrate that FADE is highly effective against state-of-the-art TBP trackers, causing significant identity switches and track terminations.

Large language models (LLMs) are increasingly deployed for everyday tasks, including food preparation and health-related guidance. However, food safety remains a high-stakes domain where inaccurate or misleading information can cause severe real-world harm. Despite these risks, current LLMs and safety guardrails lack rigorous alignment tailored to domain-specific food hazards. To address this gap, we introduce FoodGuardBench, the first comprehensive benchmark comprising 3,339 queries grounded in FDA guidelines, designed to evaluate the safety and robustness of LLMs. By constructing a taxonomy of food safety principles and employing representative jailbreak attacks (e.g., AutoDAN and PAP), we systematically evaluate existing LLMs and guardrails. Our evaluation results reveal three critical vulnerabilities: First, current LLMs exhibit sparse safety alignment in the food-related domain, easily succumbing to a few canonical jailbreak strategies. Second, when compromised, LLMs frequently generate actionable yet harmful instructions, inadvertently empowering malicious actors and posing tangible risks. Third, existing LLM-based guardrails systematically overlook these domain-specific threats, failing to detect a substantial volume of malicious inputs. To mitigate these vulnerabilities, we introduce FoodGuard-4B, a specialized guardrail model fine-tuned on our datasets to safeguard LLMs within food-related domains.

The global Lipschitz constant of a neural network governs both adversarial robustness and generalization. Conventional approaches to ``certified training" typically follow a train-then-verify paradigm: they train a network and then attempt to bound its Lipschitz constant. Because the efficient ``trivial bound" (the product of the layerwise Lipschitz constants) is exponentially loose for arbitrary networks, these approaches must rely on computationally expensive techniques such as semidefinite programming, mixed-integer programming, or branch-and-bound. We propose a different paradigm: rather than designing complex verifiers for arbitrary networks, we design networks to be verifiable by the fast trivial bound. We show that directly penalizing the trivial bound during training forces it to become tight, thereby effectively regularizing the true Lipschitz constant. To achieve this, we identify three structural obstructions to a tight trivial bound (dead neurons, bias terms, and ill-conditioned weights) and introduce architectural mitigations, including a novel notion of norm-saturating polyactivations and bias-free sinusoidal layers. Our approach avoids the runtime complexity of advanced verification while achieving strong results: we train robust networks on MNIST with Lipschitz bounds that are small (orders of magnitude lower than comparable works) and tight (within 10% of the ground truth). The experimental results validate the theoretical guarantees, support the proposed mechanisms, and extend empirically to diverse activations and non-Euclidean norms.

High-privilege LLM agents that autonomously process external documentation are increasingly trusted to automate tasks by reading and executing project instructions, yet they are granted terminal access, filesystem control, and outbound network connectivity with minimal security oversight. We identify and systematically measure a fundamental vulnerability in this trust model, which we term the \emph{Trusted Executor Dilemma}: agents execute documentation-embedded instructions, including adversarial ones, at high rates because they cannot distinguish malicious directives from legitimate setup guidance. This vulnerability is a structural consequence of the instruction-following design paradigm, not an implementation bug. To structure our measurement, we formalize a three-dimensional taxonomy covering linguistic disguise, structural obfuscation, and semantic abstraction, and construct \textbf{ReadSecBench}, a benchmark of 500 real-world README files enabling reproducible evaluation. Experiments on the commercially deployed computer-use agent show end-to-end exfiltration success rates up to 85\%, consistent across five programming languages and three injection positions. Cross-model evaluation on four LLM families in a simulation environment confirms that semantic compliance with injected instructions is consistent across model families. A 15-participant user study yields a 0\% detection rate across all participants, and evaluation of 12 rule-based and 6 LLM-based defenses shows neither category achieves reliable detection without unacceptable false-positive rates. Together, these results quantify a persistent \emph{Semantic-Safety Gap} between agents' functional compliance and their security awareness, establishing that documentation-embedded instruction injection is a persistent and currently unmitigated threat to high-privilege LLM agent deployments.

Safety alignment aims to ensure that large language models (LLMs) refuse harmful requests by post-training on harmful queries paired with refusal answers. Although safety alignment is widely adopted in industry, the overrefusal problem where aligned LLMs also reject benign queries after safety alignment post-training, remains insufficiently studied. Such an issue degrades the usability of safety alignment in real-world applications. In this paper, we examine how overrefusal arises under safety alignment, and propose a mitigation strategy inspired by our findings. We define refusal triggers as linguistic cues in the training data that elicit refusal responses, safety alignment encourages LLMs to associate refusal triggers within a training sample with refusal responses, leading aligned LLMs to refuse harmful queries. However, the refusal triggers include not only harmful linguistic cues but also non-harmful cues, therefore causing overrefusal to benign queries. Building on this mechanistic analysis, we propose a method that explicitly considers refusal triggers in the safety alignment fine-tuning. Empirical results demonstrate that our approach achieves a more favorable trade-off between defense against jailbreak attacks and responsiveness to benign queries, outperforming prior methods. Warning: this paper contains harmful and biased sentences.

AI agents that combine large language models with non-AI system components are rapidly emerging in real-world applications, offering unprecedented automation and flexibility. However, this unprecedented flexibility introduces complex security challenges fundamentally different from those in traditional software systems. This paper presents the first systematic and comprehensive survey of AI agent security, including an analysis of the design space, attack landscape, and defense mechanisms for secure AI agent systems. We further conduct case studies to point out existing gaps in securing agentic AI systems and identify open challenges in this emerging domain. Our work also introduces the first systematic framework for understanding the security risks and defense strategies of AI agents, serving as a foundation for building both secure agentic systems and advancing research in this critical area.

Cryptographic provenance standards such as C2PA and invisible watermarking are positioned as complementary defenses for content authentication, yet the two verification layers are technically independent: neither conditions on the output of the other. This work formalizes and empirically demonstrates the $\textit{Integrity Clash}$, a condition in which a digital asset carries a cryptographically valid C2PA manifest asserting human authorship while its pixels simultaneously carry a watermark identifying it as AI-generated, with both signals passing their respective verification checks in isolation. We construct metadata washing workflows that produce these authenticated fakes through standard editing pipelines, requiring no cryptographic compromise, only the semantic omission of a single assertion field permitted by the current C2PA specification. To close this gap, we propose a cross-layer audit protocol that jointly evaluates provenance metadata and watermark detection status, achieving 100% classification accuracy across 3,500 test images spanning four conflict-matrix states and three realistic perturbation conditions. Our results demonstrate that the gap between these verification layers is unnecessary and technically straightforward to close.

Modern generative agents such as OpenClaw - an open-source, self-hosted personal assistant with a community skill ecosystem, are gaining attention and are used pervasively. However, the openness and rapid growth of these ecosystems often outpace systematic security evaluation. In this paper, we design, implement, and evaluate Clawdrain, a Trojanized skill that induces a multi-turn "Segmented Verification Protocol" via injected SKILL.md instructions and a companion script that returns PROGRESS/REPAIR/TERMINAL signals. We deploy Clawdrain in a production-like OpenClaw instance with real API billing and a production model (Gemini 2.5 Pro), and we measure 6-7x token amplification over a benign baseline, with a costly, failure configuration reaching approximately 9x. We observe a deployment-only phenomenon: the agent autonomously composes general-purpose tools (e.g., shell/Python) to route around brittle protocol steps, reducing amplification and altering attack dynamics. Finally, we identify production vectors enabled by OpenClaw's architecture, including SKILL.md prompt bloat, persistent tool-output pollution, cron/heartbeat frequency amplification, and behavioral instruction injection. Overall, we demonstrate that token-drain attacks remain feasible in real deployments, but their magnitude and observability are shaped by tool composition, recovery behavior, and interface design.

The need for machine unlearning is critical for data privacy, yet existing methods often cause Knowledge Contamination by unintentionally damaging related knowledge. Such a degraded model performance after unlearning has been recently leveraged for new inference and backdoor attacks. Most studies design adversarial unlearning requests that require poisoning or duplicating training data. In this study, we introduce a new unlearning-induced attack model, namely indirect unlearning attack, which does not require data manipulation but exploits the consequence of knowledge contamination to perturb the model accuracy on security-critical predictions. To mitigate this attack, we introduce a theoretical framework that models neural networks as Neural Knowledge Systems. Based on this, we propose ROKA, a robust unlearning strategy centered on Neural Healing. Unlike conventional unlearning methods that only destroy information, ROKA constructively rebalances the model by nullifying the influence of forgotten data while strengthening its conceptual neighbors. To the best of our knowledge, our work is the first to provide a theoretical guarantee for knowledge preservation during unlearning. Evaluations on various large models, including vision transformers, multi-modal models, and large language models, show that ROKA effectively unlearns targets while preserving, or even enhancing, the accuracy of retained data, thereby mitigating the indirect unlearning attacks.

Commercial large language models are typically deployed as black-box API services, requiring users to trust providers to execute inference correctly and report token usage honestly. We present IMMACULATE, a practical auditing framework that detects economically motivated deviations-such as model substitution, quantization abuse, and token overbilling-without trusted hardware or access to model internals. IMMACULATE selectively audits a small fraction of requests using verifiable computation, achieving strong detection guarantees while amortizing cryptographic overhead. Experiments on dense and MoE models show that IMMACULATE reliably distinguishes benign and malicious executions with under 1% throughput overhead. Our code is published at https://github.com/guo-yanpei/Immaculate.

Agentic systems built on large language models (LLMs) extend beyond text generation to autonomously retrieve information and invoke tools. This runtime execution model shifts the attack surface from build-time artifacts to inference-time dependencies, exposing agents to manipulation through untrusted data and probabilistic capability resolution. While prior work has focused on model-level vulnerabilities, security risks emerging from cyclic and interdependent runtime behavior remain fragmented. We systematize these risks within a unified runtime framework, categorizing threats into data supply chain attacks (transient context injection and persistent memory poisoning) and tool supply chain attacks (discovery, implementation, and invocation). We further identify the Viral Agent Loop, in which agents act as vectors for self-propagating generative worms without exploiting code-level flaws. Finally, we advocate a Zero-Trust Runtime Architecture that treats context as untrusted control flow and constrains tool execution through cryptographic provenance rather than semantic inference.

Generative AI models pose a significant challenge to intellectual property (IP), as they can replicate unique artistic styles and concepts without attribution. While watermarking offers a potential solution, existing methods often fail in complex scenarios where multiple concepts (e.g., an object and an artistic style) are composed within a single image. These methods struggle to disentangle and attribute each concept individually. In this work, we introduce TokenTrace, a novel proactive watermarking framework for robust, multi-concept attribution. Our method embeds secret signatures into the semantic domain by simultaneously perturbing the text prompt embedding and the initial latent noise that guide the diffusion model's generation process. For retrieval, we propose a query-based TokenTrace module that takes the generated image and a textual query specifying which concepts need to be retrieved (e.g., a specific object or style) as inputs. This query-based mechanism allows the module to disentangle and independently verify the presence of multiple concepts from a single generated image. Extensive experiments show that our method achieves state-of-the-art performance on both single-concept (object and style) and multi-concept attribution tasks, significantly outperforming existing baselines while maintaining high visual quality and robustness to common transformations.

The proliferation of Large Language Models (LLMs) necessitates efficient mechanisms to distinguish machine-generated content from human text. While statistical watermarking has emerged as a promising solution, existing methods suffer from two critical limitations: the lack of a principled approach for selecting sampling distributions and the reliance on fixed-horizon hypothesis testing, which precludes valid early stopping. In this paper, we bridge this gap by developing the first e-value-based watermarking framework, Anchored E-Watermarking, that unifies optimal sampling with anytime-valid inference. Unlike traditional approaches where optional stopping invalidates Type-I error guarantees, our framework enables valid, anytime-inference by constructing a test supermartingale for the detection process. By leveraging an anchor distribution to approximate the target model, we characterize the optimal e-value with respect to the worst-case log-growth rate and derive the optimal expected stopping time. Our theoretical claims are substantiated by simulations and evaluations on established benchmarks, showing that our framework can significantly enhance sample efficiency, reducing the average token budget required for detection by 13-15% relative to state-of-the-art baselines.

Lifelong multimodal agents must continuously adapt to new tasks through post-training, but this creates fundamental tension between acquiring capabilities and preserving safety alignment. We demonstrate that fine-tuning aligned vision-language models on narrow-domain harmful datasets induces severe emergent misalignment that generalizes broadly across unrelated tasks and modalities. Through experiments on Gemma3-4B, we show that misalignment scales monotonically with LoRA rank, and that multimodal evaluation reveals substantially higher misalignment ($70.71 \pm 1.22$ at $r=128$) than text-only evaluation ($41.19 \pm 2.51$), suggesting that unimodal safety benchmarks may underestimate alignment degradation in vision-language models. Critically, even 10\% harmful data in the training mixture induces substantial alignment degradation. Geometric analysis reveals that harmful behaviors occupy a remarkably low-dimensional subspace, with the majority of misalignment information captured in 10 principal components. To mitigate misalignment, we evaluate two strategies: benign narrow fine-tuning and activation-based steering. While both approaches substantially reduce misalignment, neither completely removes the learned harmful behaviors. Our findings highlight the need for robust continual learning frameworks, as current post-training paradigms may not sufficiently preserve alignment in post-deployment settings.

LLM-based agents are becoming increasingly capable, yet their safety lags behind. This creates a gap between what agents can do and should do. This gap widens as agents engage in multi-turn interactions and employ diverse tools, introducing new risks overlooked by existing benchmarks. To systematically scale safety testing into multi-turn, tool-realistic settings, we propose a principled taxonomy that transforms single-turn harmful tasks into multi-turn attack sequences. Using this taxonomy, we construct MT-AgentRisk (Multi-Turn Agent Risk Benchmark), the first benchmark to evaluate multi-turn tool-using agent safety. Our experiments reveal substantial safety degradation: the Attack Success Rate (ASR) increases by 16% on average across open and closed models in multi-turn settings. To close this gap, we propose ToolShield, a training-free, tool-agnostic, self-exploration defense: when encountering a new tool, the agent autonomously generates test cases, executes them to observe downstream effects, and distills safety experiences for deployment. Experiments show that ToolShield effectively reduces ASR by 30% on average in multi-turn interactions. Our code is available at https://github.com/CHATS-lab/ToolShield.

Jailbreaking large language models (LLMs) has emerged as a critical security challenge with the widespread deployment of conversational AI systems. Adversarial users exploit these models through carefully crafted prompts to elicit restricted or unsafe outputs, a phenomenon commonly referred to as Jailbreaking. Despite numerous proposed defense mechanisms, attackers continue to develop adaptive prompting strategies, and existing models remain vulnerable. This motivates approaches that examine the internal behavior of LLMs rather than relying solely on prompt-level defenses. In this work, we study jailbreaking from both security and interpretability perspectives by analyzing how internal representations differ between jailbreak and benign prompts. We conduct a systematic layer-wise analysis across multiple open-source models, including GPT-J, LLaMA, Mistral, and the state-space model Mamba, and identify consistent latent-space patterns associated with harmful inputs. We then propose a tensor-based latent representation framework that captures structure in hidden activations and enables lightweight jailbreak detection without model fine-tuning or auxiliary LLM-based detectors. We further demonstrate that the latent signals can be used to actively disrupt jailbreak execution at inference time. On an abliterated LLaMA-3.1-8B model, selectively bypassing high-susceptibility layers blocks 78% of jailbreak attempts while preserving benign behavior on 94% of benign prompts. This intervention operates entirely at inference time and introduces minimal overhead, providing a scalable foundation for achieving stronger coverage by incorporating additional attack distributions or more refined susceptibility thresholds. Our results provide evidence that jailbreak behavior is rooted in identifiable internal structures and suggest a complementary, architecture-agnostic direction for improving LLM security.

Machine unlearning aims to remove the influence of a designated forget set from a trained model while preserving utility on the retained data. In modern deep networks, approximate unlearning frequently fails under large or adversarial deletions due to pronounced layer-wise heterogeneity: some layers exhibit stable, well-regularized representations while others are brittle, undertrained, or overfit, so naive update allocation can trigger catastrophic forgetting or unstable dynamics. We propose Statistical-Roughness Adaptive Gradient Unlearning (SRAGU), a mechanism-first unlearning algorithm that reallocates unlearning updates using layer-wise statistical roughness operationalized via heavy-tailed spectral diagnostics of layer weight matrices. Starting from an Adaptive Gradient Unlearning (AGU) sensitivity signal computed on the forget set, SRAGU estimates a WeightWatcher-style heavy-tailed exponent for each layer, maps it to a bounded spectral stability weight, and uses this stability signal to spectrally reweight the AGU sensitivities before applying the same minibatch update form. This concentrates unlearning motion in spectrally stable layers while damping updates in unstable or overfit layers, improving stability under hard deletions. We evaluate unlearning via behavioral alignment to a gold retrained reference model trained from scratch on the retained data, using empirical prediction-divergence and KL-to-gold proxies on a forget-focused query set; we additionally report membership inference auditing as a complementary leakage signal, treating forget-set points as should-be-forgotten members during evaluation.