Safety-aligned language models must refuse harmful requests without collapsing into broad over-refusal, but the training-time mechanisms behind this tradeoff remain unclear. Prior work characterizes refusal directions and jailbreak robustness, yet does not explain how dynamic adversarial fine-tuning changes refusal carriers across training. We present a measurement-driven mechanism study, not a new defense, on one 7B backbone under supervised fine-tuning (SFT) and R2D2-style dynamic adversarial fine-tuning. Our protocol aligns fixed-source HarmBench, StrongREJECT, and XSTest with a five-anchor refusal-geometry suite and causal interventions. R2D2 drives fixed-source HarmBench ASR to 0.000 at steps 50 and 100, then partially reopens to 0.035 at step 250 and 0.250 at step 500; SFT remains less robust, with ASR between 0.505 and 0.588 at the same anchors. On XSTest, R2D2 any-refusal is 1.000 early, then falls to 0.664 and 0.228. Geometrically, R2D2 preserves a late-layer admissible carrier through step 100 before relocating to an early-layer carrier, while effective rank remains near 1.23--1.27. Causal interventions indicate low-dimensional but utility-coupled control. These results support a reorganization account rather than a drift-only account, with evidence limited to one backbone and fixed-source attacks.

Large Language Model (LLM) agents are increasingly integrated into critical systems, leveraging external tools to interact with the real world. However, this capability exposes them to Indirect Prompt Injection (IPI), where attackers embed malicious instructions into retrieved content to manipulate the agent into executing unauthorized or unintended actions. Existing defenses predominantly focus on the pre-processing stage, neglecting the monitoring of the model's actual behavior. In this paper, we propose PlanGuard, a training-free defense framework based on the principle of Context Isolation. Unlike prior methods, PlanGuard introduces an isolated Planner that generates a reference set of valid actions derived solely from user instructions. In addition, we design a Hierarchical Verification Mechanism that first enforces strict hard constraints to block unauthorized tool invocations, and subsequently employs an Intent Verifier to validate whether parameter deviations are benign formatting variances or malicious hijacking. Experiments on the InjecAgent benchmark demonstrate that PlanGuard effectively neutralizes these attacks, reducing the Attack Success Rate (ASR) from 72.8% to 0%, while maintaining an acceptable False Positive Rate of 1.49%. Furthermore, our method is model-agnostic and highly compatible.

Recent advances in visual-language alignment have endowed vision-language models (VLMs) with fine-grained image understanding capabilities. However, this progress also introduces new privacy risks. This paper first proposes a novel privacy threat model named identity-affiliation learning: an attacker fine-tunes a VLM using only a few private photos of a target individual, thereby embedding associations between the target facial identity and their private property and social relationships into the model's internal representations. Once deployed via public APIs, this model enables unauthorized exposure of the target user's private information upon input of their photos. To benchmark VLMs' susceptibility to such identity-affiliation leakage, we introduce the first identity-affiliation dataset comprising seven typical scenarios appearing in private photos. Each scenario is instantiated with multiple identity-centered photo-description pairs. Experimental results demonstrate that mainstream VLMs like LLaVA, Qwen-VL, and MiniGPT-v2, can recognize facial identities and infer identity-affiliation relationships by fine-tuning on small-scale private photographic dataset, and even on synthetically generated datasets. To mitigate this privacy risk, we propose DP2-VL, the first Dataset Protection framework for private photos that leverages Data Poisoning. Though optimizing imperceptible perturbations by pushing the original representations toward an antithetical region, DP2-VL induces a dataset-level shift in the embedding space of VLMs'encoders. This shift separates protected images from clean inference images, causing fine-tuning on the protected set to overfit. Extensive experiments demonstrate that DP2-VL achieves strong generalization across models, robustness to diverse post-processing operations, and consistent effectiveness across varying protection ratios.

Large Language Models(LLMs) are widely deployed, yet are vulnerable to jailbreak prompts that elicit policy-violating outputs. Although prior studies have uncovered these risks, they typically treat all tokens as equally important during prompt mutation, overlooking the varying contributions of individual tokens to triggering model refusals. Consequently, these attacks introduce substantial redundant searching under query-constrained scenarios, reducing attack efficiency and hindering comprehensive vulnerability assessment. In this work, we conduct a token-level analysis of refusal behavior and observe that token contributions are highly skewed rather than uniform. Moreover, we find strong cross-model consistency in refusal tendencies, enabling the use of a surrogate model to estimate token-level contributions to the target model's refusals. Motivated by these findings, we propose TriageFuzz, a token-aware jailbreak fuzzing framework that adapts the fuzz testing approach with a series of customized designs. TriageFuzz leverages a surrogate model to estimate the contribution of individual tokens to refusal behaviors, enabling the identification of sensitive regions within the prompt. Furthermore, it incorporates a refusal-guided evolutionary strategy that adaptively weights candidate prompts with a lightweight scorer to steer the evolution toward bypassing safety constraints. Extensive experiments on six open-source LLMs and three commercial APIs demonstrate that TriageFuzz achieves comparable attack success rates (ASR) with significantly reduced query costs. Notably, it attains a 90% ASR with over 70% fewer queries compared to baselines. Even under an extremely restrictive budget of 25 queries, TriageFuzz outperforms existing methods, improving ASR by 20-40%.

Although infrared pedestrian detectors have been widely deployed in visual perception tasks, their vulnerability to physical adversarial attacks is becoming increasingly apparent. Existing physical attack methods predominantly rely on instance-specific online optimization and rigid pattern design, leading to high deployment costs and insufficient physical robustness. To address these limitations, this work proposes the Universal Physical Patch Attack (UPPA), the first universal physical attack method in the infrared domain. This method employs geometrically constrained parameterized Bezier blocks to model perturbations and utilizes the Particle Swarm Optimization (PSO) algorithm to perform unified optimization across the global data distribution, thus maintaining topological stability under dynamic deformations. In the physical deployment phase, we materialize the optimized digital perturbations into physical cold patches, achieving a continuous and smooth low-temperature distribution that naturally aligns with the thermal radiation characteristics of infrared imaging. Extensive experiments demonstrate that UPPA achieves an outstanding physical attack success rate without any online computational overhead, while also exhibiting strong cross-domain generalization and reliable black-box transferability.

Many learning systems now use graph data in which each node also contains text, such as papers with abstracts or users with posts. Because these texts often come from open platforms, an attacker may be able to quietly poison a small part of the training data and later make the model produce wrong predictions on demand. This paper studies that risk in a realistic setting where the attacker edits only node text and does not change the graph structure. We propose TAGBD, a text-only backdoor attack for text-attributed graphs. TAGBD first finds training nodes that are easier to influence, then generates natural-looking trigger text with the help of a shadow graph model, and finally injects the trigger by either replacing the original text or appending a short phrase. Experiments on three benchmark datasets show that the attack is highly effective, transfers across different graph models, and remains strong under common defenses. These results demonstrate that text alone is a practical attack channel in graph learning systems and suggest that future defenses should inspect both graph links and node content.

Although text-to-image diffusion models exhibit remarkable generative power, concept erasure techniques are essential for their safe deployment to prevent the creation of harmful content. This has fostered a dynamic interplay between the development of erasure defenses and the adversarial probes designed to bypass them, and this co-evolution has progressively enhanced the efficacy of erasure methods. However, this adversarial co-evolution has converged on a narrow, text-centric paradigm that equates erasure with severing the text-to-image mapping, ignoring that the underlying visual knowledge related to undesired concepts still persist. To substantiate this claim, we investigate from a visual perspective, leveraging DDIM inversion to probe whether a generative pathway for the erased concept can still be found. However, identifying such a visual generative pathway is challenging because standard text-guided DDIM inversion is actively resisted by text-centric defenses within the erased model. To address this, we introduce TINA, a novel Text-free INversion Attack, which enforces this visual-only probe by operating under a null-text condition, thereby avoiding existing text-centric defenses. Moreover, TINA integrates an optimization procedure to overcome the accumulating approximation errors that arise when standard inversion operates without its usual textual guidance. Our experiments demonstrate that TINA regenerates erased concepts from models treated with state-of-the-art unlearning. The success of TINA proves that current methods merely obscure concepts, highlighting an urgent need for paradigms that operate directly on internal visual knowledge.

To generalize deepfake detectors to future unseen forgeries, most existing methods attempt to simulate the dynamically evolving forgery types using available source domain data. However, predicting an unbounded set of future manipulations from limited prior examples is infeasible. To overcome this limitation, we propose to exploit the invariance of \textbf{real data} from two complementary perspectives: the fixed population distribution of the entire real class and the inherent Gaussianity of individual real images. Building on these properties, we introduce the Real Distribution Bias Correction (RDBC) framework, which consists of two key components: the Real Population Distribution Estimation module and the Distribution-Sampled Feature Whitening module. The former utilizes the independent and identically distributed (\iid) property of real samples to derive the normal distribution form of their statistics, from which the distribution parameters can be estimated using limited source domain data. Based on the learned population distribution, the latter utilizes the inherent Gaussianity of real data as a discriminative prior and performs a sampling-based whitening operation to amplify the Gaussianity gap between real and fake samples. Through synergistic coupling of the two modules, our model captures the real-world properties of real samples, thereby enhancing its generalizability to unseen target domains. Extensive experiments demonstrate that RDBC achieves state-of-the-art performance in both in-domain and cross-domain deepfake detection.

Code agents powered by large language models can execute shell commands on behalf of users, introducing severe security vulnerabilities. This paper presents a two-phase security analysis of the OpenClaw platform. As an open-source AI agent framework that operates locally, OpenClaw can be integrated with various commercial large language models. Because its native architecture lacks built-in security constraints, it serves as an ideal subject for evaluating baseline agent vulnerabilities. First, we systematically evaluate OpenClaw's native resilience against malicious instructions. By testing 47 adversarial scenarios across six major attack categories derived from the MITRE ATLAS and ATT\&CK frameworks, we have demonstrated that OpenClaw exhibits significant inherent security issues. It primarily relies on the security capabilities of the backend LLM and is highly susceptible to sandbox escape attacks, with an average defense rate of only 17\%. To mitigate these critical security gaps, we propose and implement a novel Human-in-the-Loop (HITL) defense layer. We utilize a dual-mode testing framework to evaluate the system with and without our proposed intervention. Our findings show that the introduced HITL layer significantly hardens the system, successfully intercepting up to 8 severe attacks that completely bypassed OpenClaw's native defenses. By combining native capabilities with our HITL approach, the overall defense rate improves to a range of 19\% to 92\%. Our study not only exposes the intrinsic limitations of current code agents but also demonstrates the effectiveness of human-agent collaborative defense strategies.

The Model Context Protocol (MCP) is an open and standardized interface that enables large language models (LLMs) to interact with external tools and services, and is increasingly adopted by AI agents. However, the security of MCP-based systems remains largely unexplored.In this work, we conduct a large-scale security analysis of MCP servers integrated within MCP clients. We show that treating MCP servers as trusted entities without authenticating the caller identity is fundamentally insecure. Since MCP servers often cannot distinguish who is invoking a request, a single authorization decision may implicitly grant access to multiple, potentially untrusted callers.Our empirical study reveals that most MCP servers rely on persistent authorization states, allowing tool invocations after an initial authorization without re-authentication, regardless of the caller. In addition, many MCP servers fail to enforce authentication at the per-tool level, enabling unauthorized access to sensitive operations.These findings demonstrate that one-time authorization and server-level trust significantly expand the attack surface of MCP-based systems, highlighting the need for explicit caller authentication and fine-grained authorization mechanisms.

LLM agents often rely on Skills to describe available tools and recommended procedures. We study a hidden-comment prompt injection risk in this documentation layer: when a Markdown Skill is rendered to HTML, HTML comment blocks can become invisible to human reviewers, yet the raw text may still be supplied verbatim to the model. In experiments, we find that DeepSeek-V3.2 and GLM-4.5-Air can be influenced by malicious instructions embedded in a hidden comment appended to an otherwise legitimate Skill, yielding outputs that contain sensitive tool intentions. A short defensive system prompt that treats Skills as untrusted and forbids sensitive actions prevents these malicious tool calls and instead surfaces the suspicious hidden instructions.

The Model Context Protocol (MCP) enables large language models to invoke external tools through natural-language descriptions, forming the foundation of many AI agent applications. However, MCP does not enforce consistency between documented tool behavior and actual code execution, even though MCP Servers often run with broad system privileges. This gap introduces a largely unexplored security risk. We study how mismatches between externally presented tool descriptions and underlying implementations systematically shape the mental models and decision-making behavior of intelligent agents. Specifically, we present the first large-scale study of description-code inconsistency in the MCP ecosystem. We design an automated static analysis framework and apply it to 10,240 real-world MCP Servers across 36 categories. Our results show that while most servers are highly consistent, approximately 13% exhibit substantial mismatches that can enable undocumented privileged operations, hidden state mutations, or unauthorized financial actions. We further observe systematic differences across application categories, popularity levels, and MCP marketplaces. Our findings demonstrate that description-code inconsistency is a concrete and prevalent attack surface in MCP-based AI agents, and motivate the need for systematic auditing and stronger transparency guarantees in future agent ecosystems.

In-context learning (ICL) has become a powerful, data-efficient paradigm for text classification using large language models. However, its robustness against realistic adversarial threats remains largely unexplored. We introduce ICL-Evader, a novel black-box evasion attack framework that operates under a highly practical zero-query threat model, requiring no access to model parameters, gradients, or query-based feedback during attack generation. We design three novel attacks, Fake Claim, Template, and Needle-in-a-Haystack, that exploit inherent limitations of LLMs in processing in-context prompts. Evaluated across sentiment analysis, toxicity, and illicit promotion tasks, our attacks significantly degrade classifier performance (e.g., achieving up to 95.3% attack success rate), drastically outperforming traditional NLP attacks which prove ineffective under the same constraints. To counter these vulnerabilities, we systematically investigate defense strategies and identify a joint defense recipe that effectively mitigates all attacks with minimal utility loss (<5% accuracy degradation). Finally, we translate our defensive insights into an automated tool that proactively fortifies standard ICL prompts against adversarial evasion. This work provides a comprehensive security assessment of ICL, revealing critical vulnerabilities and offering practical solutions for building more robust systems. Our source code and evaluation datasets are publicly available at: https://github.com/ChaseSecurity/ICL-Evader .

Fine-Tuning-as-a-Service (FTaaS) facilitates the customization of Multimodal Large Language Models (MLLMs) but introduces critical backdoor risks via poisoned data. Existing defenses either rely on supervised signals or fail to generalize across diverse trigger types and modalities. In this work, we uncover a universal backdoor fingerprint-attention allocation divergence-where poisoned samples disrupt the balanced attention distribution across three functional components: system instructions, vision inputs, and user textual queries, regardless of trigger morphology. Motivated by this insight, we propose Tri-Component Attention Profiling (TCAP), an unsupervised defense framework to filter backdoor samples. TCAP decomposes cross-modal attention maps into the three components, identifies trigger-responsive attention heads via Gaussian Mixture Model (GMM) statistical profiling, and isolates poisoned samples through EM-based vote aggregation. Extensive experiments across diverse MLLM architectures and attack methods demonstrate that TCAP achieves consistently strong performance, establishing it as a robust and practical backdoor defense in MLLMs.

Federated graph learning (FGL) has recently emerged as a promising privacy-preserving paradigm that enables distributed graph learning across multiple data owners. A critical privacy concern in federated learning is whether an adversary can recover raw data from shared gradients, a vulnerability known as deep leakage from gradients (DLG). However, most prior studies on the DLG problem focused on image or text data, and it remains an open question whether graphs can be effectively recovered, particularly when the graph structure and node features are uniquely entangled in GNNs. In this work, we first theoretically analyze the components in FGL and derive a crucial insight: once the graph structure is recovered, node features can be obtained through a closed-form recursive rule. Building on this analysis, we propose GraphDLG, a novel approach to recover raw training graphs from shared gradients in FGL, which can utilize randomly generated graphs or client-side training graphs as auxiliaries to enhance recovery. Extensive experiments demonstrate that GraphDLG outperforms existing solutions by successfully decoupling the graph structure and node features, achieving improvements of over 5.46% (by MSE) for node feature reconstruction and over 25.04% (by AUC) for graph structure reconstruction.

The proliferation of powerful Text-to-Video (T2V) models, trained on massive web-scale datasets, raises urgent concerns about copyright and privacy violations. Membership inference attacks (MIAs) provide a principled tool for auditing such risks, yet existing techniques - designed for static data like images or text - fail to capture the spatio-temporal complexities of video generation. In particular, they overlook the sparsity of memorization signals in keyframes and the instability introduced by stochastic temporal dynamics. In this paper, we conduct the first systematic study of MIAs against T2V models and introduce a novel framework VidLeaks, which probes sparse-temporal memorization through two complementary signals: 1) Spatial Reconstruction Fidelity (SRF), using a Top-K similarity to amplify spatial memorization signals from sparsely memorized keyframes, and 2) Temporal Generative Stability (TGS), which measures semantic consistency across multiple queries to capture temporal leakage. We evaluate VidLeaks under three progressively restrictive black-box settings - supervised, reference-based, and query-only. Experiments on three representative T2V models reveal severe vulnerabilities: VidLeaks achieves AUC of 82.92% on AnimateDiff and 97.01% on InstructVideo even in the strict query-only setting, posing a realistic and exploitable privacy risk. Our work provides the first concrete evidence that T2V models leak substantial membership information through both sparse and temporal memorization, establishing a foundation for auditing video generation systems and motivating the development of new defenses. Code is available at: https://zenodo.org/records/17972831.

Low-Rank Adaptation (LoRA) has become a popular solution for fine-tuning large language models (LLMs) in federated settings, dramatically reducing update costs by introducing trainable low-rank matrices. However, when integrated with frameworks like FedIT, LoRA introduces a critical vulnerability: clients submit $A$ and $B$ matrices separately, while only their product $AB$ determines the model update, yet this composite is never directly verified. We propose Gradient Assembly Poisoning (GAP), a novel attack that exploits this blind spot by crafting individually benign $A$ and $B$ matrices whose product yields malicious updates. GAP operates without access to training data or inter-client coordination and remains undetected by standard anomaly detectors. We identify four systemic vulnerabilities in LoRA-based federated systems and validate GAP across LLaMA, ChatGLM, and GPT-2. GAP consistently induces degraded or biased outputs while preserving surface fluency, reducing BLEU by up to 14.5\%, increasing factual and grammatical errors by over 800\%, and maintaining 92.6\% long-form response length. These results reveal a new class of stealthy, persistent threats in distributed LoRA fine-tuning.

Pseudorandom error-correcting codes (PRC) is a novel cryptographic primitive proposed at CRYPTO 2024. Due to the dual capability of pseudorandomness and error correction, PRC has been recognized as a promising foundational component for watermarking AI-generated content. However, the security of PRC has not been thoroughly analyzed, especially with concrete parameters or even in the face of cryptographic attacks. To fill this gap, we present the first cryptanalysis of PRC. We first propose three attacks to challenge the undetectability and robustness assumptions of PRC. Among them, two attacks aim to distinguish PRC-based codewords from plain vectors, and one attack aims to compromise the decoding process of PRC. Our attacks successfully undermine the claimed security guarantees across all parameter configurations. Notably, our attack can detect the presence of a watermark with overwhelming probability at a cost of $2^{22}$ operations. We also validate our approach by attacking real-world large generative models such as DeepSeek and Stable Diffusion. To mitigate our attacks, we further propose three defenses to enhance the security of PRC, including parameter suggestions, implementation suggestions, and constructing a revised key generation algorithm. Our proposed revised key generation function effectively prevents the occurrence of weak keys. However, we highlight that the current PRC-based watermarking scheme still cannot achieve a 128-bit security under our parameter suggestions due to the inherent configurations of large generative models, such as the maximum output length of large language models.

The rapid proliferation of Artificial Intelligence Generated Content has precipitated a crisis of trust and urgent regulatory demands. However, existing identification tools suffer from fragmentation and a lack of support for visible compliance marking. To address these gaps, we introduce the \textbf{UniMark}, an open-source, unified framework for multimodal content governance. Our system features a modular unified engine that abstracts complexities across text, image, audio, and video modalities. Crucially, we propose a novel dual-operation strategy, natively supporting both \emph{Hidden Watermarking} for copyright protection and \emph{Visible Marking} for regulatory compliance. Furthermore, we establish a standardized evaluation framework with three specialized benchmarks (Image/Video/Audio-Bench) to ensure rigorous performance assessment. This toolkit bridges the gap between advanced algorithms and engineering implementation, fostering a more transparent and secure digital ecosystem.

Embodied AI systems integrate language models with real world sensing, mobility, and cloud connected mobile apps. Yet while model jailbreaks have drawn significant attention, the broader system stack of embodied intelligence remains largely unexplored. In this work, we conduct the first holistic security analysis of the Unitree Go2 platform and uncover ten cross layer vulnerabilities the "Ten Sins of Embodied AI Security." Using BLE sniffing, traffic interception, APK reverse engineering, cloud API testing, and hardware probing, we identify systemic weaknesses across three architectural layers: wireless provisioning, core modules, and external interfaces. These include hard coded keys, predictable handshake tokens, WiFi credential leakage, missing TLS validation, static SSH password, multilingual safety bypass behavior, insecure local relay channels, weak binding logic, and unrestricted firmware access. Together, they allow adversaries to hijack devices, inject arbitrary commands, extract sensitive information, or gain full physical control.Our findings show that securing embodied AI requires far more than aligning the model itself. We conclude with system level lessons learned and recommendations for building embodied platforms that remain robust across their entire software hardware ecosystem.