Mechanistic interpretability reveals that safety-critical behaviors (e.g., alignment, jailbreak, backdoor) in Large Language Models (LLMs) are grounded in specialized functional components. However, existing safety attribution methods struggle with generalization and reliability due to their reliance on heuristic, domain-specific metrics and search algorithms. To address this, we propose \ourmethod, a unified safety interpretability framework that identifies functionally complete safety circuits in LLMs via optimization. Unlike methods focusing on isolated heads or neurons, \ourmethod introduces differentiable binary masks to extract multi-granular circuits through gradient descent on safety datasets, while integrates Safety Circuit Tuning to utilize these sparse circuits for efficient safety fine-tuning. We validate \ourmethod in two key scenarios in LLM safety: \textbf{(1) backdoor attacks}, identifying a backdoor circuit with 0.42\% sparsity, whose ablation eradicates the Attack Success Rate (ASR) from 100\% $\to$ 0.4\% while retaining over 99\% general utility; \textbf{(2) safety alignment}, localizing an alignment circuit with 3.03\% heads and 0.79\% neurons, whose removal spikes ASR from 0.8\% $\to$ 96.9\%, whereas excluding this circuit during helpfulness fine-tuning maintains 96.5\% safety retention.

Speech deepfake source verification systems aims to determine whether two synthetic speech utterances originate from the same source generator, often assuming that the resulting source embeddings are independent of speaker traits. However, this assumption remains unverified. In this paper, we first investigate the impact of speaker factors on source verification. We propose a speaker-disentangled metric learning (SDML) framework incorporating two novel loss functions. The first leverages Chebyshev polynomial to mitigate gradient instability during disentanglement optimization. The second projects source and speaker embeddings into hyperbolic space, leveraging Riemannian metric distances to reduce speaker information and learn more discriminative source features. Experimental results on MLAAD benchmark, evaluated under four newly proposed protocols designed for source-speaker disentanglement scenarios, demonstrate the effectiveness of SDML framework. The code, evaluation protocols and demo website are available at https://github.com/xxuan-acoustics/RiemannSD-Net.

As vision-language models (VLMs) are increasingly deployed in open-world scenarios, they can be easily induced by visual jailbreak attacks to generate harmful content, posing serious risks to model safety and trustworthy usage. Recent activation steering methods inject directional vectors into model activations during inference to induce refusal behaviors and have demonstrated effectiveness. However, a steering vector may both enhance refusal ability and cause over-refusal, thereby degrading model performance on benign inputs. Moreover, due to the lack of theoretical interpretability, these methods still suffer from limited robustness and effectiveness. To better balance safety and utility, we propose NullSteer, a null-space projected activation defense framework. Our method constructs refusal directions within model activations through a linear transformation: it maintains zero perturbation within the benign subspace while dynamically inducing refusal along potentially harmful directions, thereby theoretically achieving safety enhancement without impairing the model's general capabilities. Extensive experiments show that NullSteer significantly reduces harmful outputs under various jailbreak attacks (average ASR reduction over 15 percent on MiniGPT-4) while maintaining comparable performance to the original model on general benchmarks.

Hardware faults, specifically bit-flips in quantized weights, pose a severe reliability threat to Large Language Models (LLMs), often triggering catastrophic model collapses. We demonstrate that this vulnerability fundamentally stems from the spatial alignment between sensitive weight bits and extreme activation outliers, which causes a single hardware fault to be massively amplified. To address this, we propose Rotated Robustness (RoR), a training-free defense utilizing orthogonal Householder transformations. By applying an orthogonal rotation to the activation space, RoR geometrically smooths extreme outliers across all feature dimensions. This mechanism effectively breaks the alignment between outliers and vulnerable weights, mathematically guaranteeing original model accuracy. Extensive empirical evaluations across Llama-2/3, OPT, and Qwen families demonstrate the superior reliability of our approach. Under random bit-flip attacks, RoR reduces the stochastic collapse rate from 3.15\% to 0.00\% on Qwen2.5-7B. Furthermore, under severe targeted attacks with 50 Progressive Bit Search flips, RoR sustains robust reasoning on Llama-2-7B, maintaining a 43.9\% MMLU accuracy that nearly matches its 45.2\% unattacked accuracy, while competing defenses collapse to random guessing. Most notably, against the Single-Point Fault Attack (SPFA) -- the most aggressive targeted threat -- RoR exponentially inflates the attack complexity from a few bits to over 17,000 precise bit-flips. With a negligible storage overhead of 0.31\% and a minimal inference latency increase of 9.1\% on Llama-2-7B, RoR achieves true lossless robustness, providing a practical and highly reliable defense for LLM deployment.

Graph Neural Networks (GNNs) are highly vulnerable to adversarial perturbations in both topology and features, making the learning of robust representations a critical challenge. In this work, we bridge GNNs with control theory to introduce a novel defense framework grounded in integer- and fractional-order Lyapunov stability. Unlike conventional strategies that rely on resource-heavy adversarial training or data purification, our approach fundamentally constrains the underlying feature-update dynamics of the GNN. We propose an adaptive, learnable Lyapunov function paired with a novel projection mechanism that maps the network's state into a stable space, thereby offering theoretically provable stability guarantees. Notably, this mechanism is orthogonal to existing defenses, allowing for seamless integration with techniques like adversarial training to achieve cumulative robustness. Extensive experiments demonstrate that our Lyapunov-stable graph neural flows substantially outperform base neural flows and state-of-the-art baselines across standard benchmarks and various adversarial attack scenarios.

Recent advancements in video generation technologies have been significant, resulting in their widespread application across multiple domains. However, concerns have been mounting over the potential misuse of generated content. Tracing the origin of generated videos has become crucial to mitigate potential misuse and identify responsible parties. Existing video attribution methods require additional operations or the training of source attribution models, which may degrade video quality or necessitate large amounts of training samples. To address these challenges, we define for the first time the "few-shot training-free generated video attribution" task and propose SWIFT, which is tightly integrated with the temporal characteristics of the video. By leveraging the "Pixel Frames(many) to Latent Frame(one)" temporal mapping within each video chunk, SWIFT applies a fixed-length sliding window to perform two distinct reconstructions: normal and corrupted. The variation in the losses between two reconstructions is then used as an attribution signal. We conducted an extensive evaluation of five state-of-the-art (SOTA) video generation models. Experimental results show that SWIFT achieves over 90% average attribution accuracy with merely 20 video samples across all models and even enables zero-shot attribution for HunyuanVideo, EasyAnimate, and Wan2.2. Our source code is available at https://github.com/wangchao0708/SWIFT.

Large vision-language models (LVLMs) have achieved remarkable progress in vision-language reasoning tasks, yet ensuring their safety remains a critical challenge. Recent input-side defenses detect unsafe images with CLIP and prepend safety prefixes to prompts, but they still suffer from inaccurate detection in complex scenes and unstable safety signals during decoding. To address these issues, we propose GuardAlign, a training-free defense framework that integrates two strategies. First, OT-enhanced safety detection leverages optimal transport to measure distribution distances between image patches and unsafe semantics, enabling accurate identification of malicious regions without additional computational cost. Second, cross-modal attentive calibration strengthens the influence of safety prefixes by adaptively reallocating attention across layers, ensuring that safety signals remain consistently activated throughout generation. Extensive evaluations on six representative MLLMs demonstrate that GuardAlign reduces unsafe response rates by up to 39% on SPA-VL, while preserving utility, achieving an improvement on VQAv2 from 78.51% to 79.21%.

The rise of text-to-video generation models has raised growing concerns over content authenticity, copyright protection, and malicious misuse. Watermarking serves as an effective mechanism for regulating such AI-generated content, where high fidelity and strong robustness are particularly critical. Recent generative image watermarking methods provide a promising foundation by leveraging watermark information and pseudo-random keys to control the initial sampling noise, enabling lossless embedding. However, directly extending these techniques to videos introduces two key limitations: Existing designs implicitly rely on strict alignment between video frames and frame-dependent pseudo-random binary sequences used for watermark encryption. Once this alignment is disrupted, subsequent watermark extraction becomes unreliable; and Video-specific distortions, such as inter-frame compression, significantly degrade watermark reliability. To address these issues, we propose SKeDA, a generative watermarking framework tailored for text-to-video diffusion models. SKeDA consists of two components: (1) Shuffle-Key-based Distribution-preserving Sampling (SKe) employs a single base pseudo-random binary sequence for watermark encryption and derives frame-level encryption sequences through permutation. This design transforms watermark extraction from synchronization-sensitive sequence decoding into permutation-tolerant set-level aggregation, substantially improving robustness against frame reordering and loss; and (2) Differential Attention (DA), which computes inter-frame differences and dynamically adjusts attention weights during extraction, enhancing robustness against temporal distortions. Extensive experiments demonstrate that SKeDA preserves high video generation quality and watermark robustness.

Omni-modal Large Language Models (OLLMs) greatly expand LLMs' multimodal capabilities but also introduce cross-modal safety risks. However, a systematic understanding of vulnerabilities in omni-modal interactions remains lacking. To bridge this gap, we establish a modality-semantics decoupling principle and construct the AdvBench-Omni dataset, which reveals a significant vulnerability in OLLMs. Mechanistic analysis uncovers a Mid-layer Dissolution phenomenon driven by refusal vector magnitude shrinkage, alongside the existence of a modal-invariant pure refusal direction. Inspired by these insights, we extract a golden refusal vector using Singular Value Decomposition and propose OmniSteer, which utilizes lightweight adapters to modulate intervention intensity adaptively. Extensive experiments show that our method not only increases the Refusal Success Rate against harmful inputs from 69.9% to 91.2%, but also effectively preserves the general capabilities across all modalities. Our code is available at: https://github.com/zhrli324/omni-safety-research.

Unlearnable examples (UE) have emerged as a practical mechanism to prevent unauthorized model training on private vision data, while extending this protection to tabular data is nontrivial. Tabular data in finance and healthcare is highly sensitive, yet existing UE methods transfer poorly because tabular features mix numerical and categorical constraints and exhibit saliency sparsity, with learning dominated by a few dimensions. Under a Spectral Dominance condition, we show certified unlearnability is feasible when the poison spectrum overwhelms the clean semantic spectrum. Guided by this, we propose Unlearnable Tabular Data via DecOuPled Shortcut EmbeddIng (UTOPIA), which exploits feature redundancy to decouple optimization into two channels: high saliency features for semantic obfuscation and low saliency redundant features for embedding a hyper correlated shortcut, yielding constraint-aware dominant shortcuts while preserving tabular validity. Extensive experiments across tabular datasets and models show UTOPIA drives unauthorized training toward near random performance, outperforming strong UE baselines and transferring well across architectures.

The emergence of Large Reasoning Models (LRMs) introduces a new paradigm of explicit reasoning, enabling remarkable advances yet posing unique risks such as reasoning manipulation and information leakage. To mitigate these risks, current alignment strategies predominantly rely on heavy post-training paradigms or external interventions. However, these approaches are often computationally intensive and fail to address the inherent awareness-compliance gap, a critical misalignment where models recognize potential risks yet prioritize following user instructions due to their sycophantic tendencies. To address these limitations, we propose Self-Guard, a lightweight safety defense framework that reinforces safety compliance at the representational level. Self-Guard operates through two principal stages: (1) safety-oriented prompting, which activates the model's latent safety awareness to evoke spontaneous reflection, and (2) safety activation steering, which extracts the resulting directional shift in the hidden state space and amplifies it to ensure that safety compliance prevails over sycophancy during inference. Experiments demonstrate that Self-Guard effectively bridges the awareness-compliance gap, achieving robust safety performance without compromising model utility. Furthermore, Self-Guard exhibits strong generalization across diverse unseen risks and varying model scales, offering a cost-efficient solution for LRM safety alignment.

Emergent Misalignment refers to a failure mode in which fine-tuning large language models (LLMs) on narrowly scoped data induces broadly misaligned behavior. Prior explanations mainly attribute this phenomenon to the generalization of erroneous or unsafe content. In this work, we show that this view is incomplete. Across multiple domains and model families, we find that fine-tuning models on data exhibiting specific character-level dispositions induces substantially stronger and more transferable misalignment than incorrect-advice fine-tuning, while largely preserving general capabilities. This indicates that emergent misalignment arises from stable shifts in model behavior rather than from capability degradation or corrupted knowledge. We further show that such behavioral dispositions can be conditionally activated by both training-time triggers and inference-time persona-aligned prompts, revealing shared structure across emergent misalignment, backdoor activation, and jailbreak susceptibility. Overall, our results identify character formation as a central and underexplored alignment risk, suggesting that robust alignment must address behavioral dispositions rather than isolated errors or prompt-level defenses.

Transferable adversarial attacks on point clouds remain challenging, as existing methods often rely on model-specific gradients or heuristics that limit generalization to unseen architectures. In this paper, we rethink adversarial transferability from a compact subspace perspective and propose CoSA, a transferable attack framework that operates within a shared low-dimensional semantic space. Specifically, each point cloud is represented as a compact combination of class-specific prototypes that capture shared semantic structure, while adversarial perturbations are optimized within a low-rank subspace to induce coherent and architecture-agnostic variations. This design suppresses model-dependent noise and constrains perturbations to semantically meaningful directions, thereby improving cross-model transferability without relying on surrogate-specific artifacts. Extensive experiments on multiple datasets and network architectures demonstrate that CoSA consistently outperforms state-of-the-art transferable attacks, while maintaining competitive imperceptibility and robustness under common defense strategies. Codes will be made public upon paper acceptance.

Digital watermarking is essential for securing generated images from diffusion models. Accurate watermark evaluation is critical for algorithm development, yet existing methods have significant limitations: they lack a unified framework for both residual and semantic watermarks, provide results without interpretability, neglect comprehensive security considerations, and often use inappropriate metrics for semantic watermarks. To address these gaps, we propose WMVLM, the first unified and interpretable evaluation framework for diffusion model image watermarking via vision-language models (VLMs). We redefine quality and security metrics for each watermark type: residual watermarks are evaluated by artifact strength and erasure resistance, while semantic watermarks are assessed through latent distribution shifts. Moreover, we introduce a three-stage training strategy to progressively enable the model to achieve classification, scoring, and interpretable text generation. Experiments show WMVLM outperforms state-of-the-art VLMs with strong generalization across datasets, diffusion models, and watermarking methods.

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 .

Latent-based watermarks, integrated into the generation process of latent diffusion models (LDMs), simplify detection and attribution of generated images. However, recent black-box forgery attacks, where an attacker needs at least one watermarked image and black-box access to the provider's model, can embed the provider's watermark into images not produced by the provider, posing outsized risk to provenance and trust. We propose SemBind, the first defense framework for latent-based watermarks that resists black-box forgery by binding latent signals to image semantics via a learned semantic masker. Trained with contrastive learning, the masker yields near-invariant codes for the same prompt and near-orthogonal codes across prompts; these codes are reshaped and permuted to modulate the target latent before any standard latent-based watermark. SemBind is generally compatible with existing latent-based watermarking schemes and keeps image quality essentially unchanged, while a simple mask-ratio parameter offers a tunable trade-off between anti-forgery strength and robustness. Across four mainstream latent-based watermark methods, our SemBind-enabled anti-forgery variants markedly reduce false acceptance under black-box forgery while providing a controllable robustness-security balance.

Vision-language models (VLMs) such as CLIP have demonstrated remarkable zero-shot generalization, yet remain highly vulnerable to adversarial examples (AEs). While test-time defenses are promising, existing methods fail to provide sufficient robustness against strong attacks and are often hampered by high inference latency and task-specific applicability. To address these limitations, we start by investigating the intrinsic properties of AEs, which reveals that AEs exhibit severe feature inconsistency under progressive frequency attenuation. We further attribute this to the model's inherent spectral bias. Leveraging this insight, we propose an efficient test-time defense named Contrastive Spectral Rectification (CSR). CSR optimizes a rectification perturbation to realign the input with the natural manifold under a spectral-guided contrastive objective, which is applied input-adaptively. Extensive experiments across 16 classification benchmarks demonstrate that CSR outperforms the SOTA by an average of 18.1% against strong AutoAttack with modest inference overhead. Furthermore, CSR exhibits broad applicability across diverse visual tasks. Code is available at https://github.com/Summu77/CSR.

Large language models (LLMs) are increasingly deployed over knowledge bases for efficient knowledge retrieval and question answering. However, LLMs can inadvertently answer beyond a user's permission scope, leaking sensitive content, thus making it difficult to deploy knowledge-base QA under fine-grained access control requirements. In this work, we identify a geometric regularity in intermediate activations: for the same query, representations induced by different permission scopes cluster distinctly and are readily separable. Building on this separability, we propose Activation-space Anchored Access Control (AAAC), a training-free framework for multi-class permission control. AAAC constructs an anchor bank, with one permission anchor per class, from a small offline sample set and requires no fine-tuning. At inference time, a multi-anchor steering mechanism redirects each query's activations toward the anchor-defined authorized region associated with the current user, thereby suppressing over-privileged generations by design. Finally, extensive experiments across three LLM families demonstrate that AAAC reduces permission violation rates by up to 86.5% and prompt-based attack success rates by 90.7%, while improving response usability with minor inference overhead compared to baselines.

Recent advances in generative modeling can create remarkably realistic synthetic videos, making it increasingly difficult for humans to distinguish them from real ones and necessitating reliable detection methods. However, two key limitations hinder the development of this field. \textbf{From the dataset perspective}, existing datasets are often limited in scale and constructed using outdated or narrowly scoped generative models, making it difficult to capture the diversity and rapid evolution of modern generative techniques. Moreover, the dataset construction process frequently prioritizes quantity over quality, neglecting essential aspects such as semantic diversity, scenario coverage, and technological representativeness. \textbf{From the benchmark perspective}, current benchmarks largely remain at the stage of dataset creation, leaving many fundamental issues and in-depth analysis yet to be systematically explored. Addressing this gap, we propose AIGVDBench, a benchmark designed to be comprehensive and representative, covering \textbf{31} state-of-the-art generation models and over \textbf{440,000} videos. By executing more than \textbf{1,500} evaluations on \textbf{33} existing detectors belonging to four distinct categories. This work presents \textbf{8 in-depth analyses} from multiple perspectives and identifies \textbf{4 novel findings} that offer valuable insights for future research. We hope this work provides a solid foundation for advancing the field of AI-generated video detection. Our benchmark is open-sourced at https://github.com/LongMa-2025/AIGVDBench.

To standardize interactions between LLM-based agents and their environments, the Model Context Protocol (MCP) was proposed and has since been widely adopted. However, integrating external tools expands the attack surface, exposing agents to tool poisoning attacks. In such attacks, malicious instructions embedded in tool metadata are injected into the agent context during MCP registration phase, thereby manipulating agent behavior. Prior work primarily focuses on explicit tool poisoning or relied on manually crafted poisoned tools. In contrast, we focus on a particularly stealthy variant: implicit tool poisoning, where the poisoned tool itself remains uninvoked. Instead, the instructions embedded in the tool metadata induce the agent to invoke a legitimate but high-privilege tool to perform malicious operations. We propose MCP-ITP, the first automated and adaptive framework for implicit tool poisoning within the MCP ecosystem. MCP-ITP formulates poisoned tool generation as a black-box optimization problem and employs an iterative optimization strategy that leverages feedback from both an evaluation LLM and a detection LLM to maximize Attack Success Rate (ASR) while evading current detection mechanisms. Experimental results on the MCPTox dataset across 12 LLM agents demonstrate that MCP-ITP consistently outperforms the manually crafted baseline, achieving up to 84.2% ASR while suppressing the Malicious Tool Detection Rate (MDR) to as low as 0.3%.