Vision-language-action (VLA) models have shown strong performance in robotic manipulation, yet their robustness to physically realizable adversarial attacks remains underexplored. Existing studies reveal vulnerabilities through language perturbations and 2D visual attacks, but these attack surfaces are either less representative of real deployment or limited in physical realism. In contrast, adversarial 3D textures pose a more physically plausible and damaging threat, as they are naturally attached to manipulated objects and are easier to deploy in physical environments. Bringing adversarial 3D textures to VLA systems is nevertheless nontrivial. A central obstacle is that standard 3D simulators do not provide a differentiable optimization path from the VLA objective function back to object appearance, making it difficult to optimize through an end-to-end manner. To address this, we introduce Foreground-Background Decoupling (FBD), which enables differentiable texture optimization through dual-renderer alignment while preserving the original simulation environment. To further ensure that the attack remains effective across long-horizon and diverse viewpoints in the physical world, we propose Trajectory-Aware Adversarial Optimization (TAAO), which prioritizes behaviorally critical frames and stabilizes optimization with a vertex-based parameterization. Built on these designs, we present Tex3D, the first framework for end-to-end optimization of 3D adversarial textures directly within the VLA simulation environment. Experiments in both simulation and real-robot settings show that Tex3D significantly degrades VLA performance across multiple manipulation tasks, achieving task failure rates of up to 96.7\%. Our empirical results expose critical vulnerabilities of VLA systems to physically grounded 3D adversarial attacks and highlight the need for robustness-aware training.

Modern LLMs employ safety mechanisms that extend beyond surface-level input filtering to latent semantic representations and generation-time reasoning, enabling them to recover obfuscated malicious intent during inference and refuse accordingly, and rendering many surface-level obfuscation jailbreak attacks ineffective. We propose Structured Semantic Cloaking (S2C), a novel multi-dimensional jailbreak attack framework that manipulates how malicious semantic intent is reconstructed during model inference. S2C strategically distributes and reshapes semantic cues such that full intent consolidation requires multi-step inference and long-range co-reference resolution within deeper latent representations. The framework comprises three complementary mechanisms: (1) Contextual Reframing, which embeds the request within a plausible high-stakes scenario to bias the model toward compliance; (2) Content Fragmentation, which disperses the semantic signature of the request across disjoint prompt segments; and (3) Clue-Guided Camouflage, which disguises residual semantic cues while embedding recoverable markers that guide output generation. By delaying and restructuring semantic consolidation, S2C degrades safety triggers that depend on coherent or explicitly reconstructed malicious intent at decoding time, while preserving sufficient instruction recoverability for functional output generation. We evaluate S2C across multiple open-source and proprietary LLMs using HarmBench and JBB-Behaviors, where it improves Attack Success Rate (ASR) by 12.4% and 9.7%, respectively, over the current SOTA. Notably, S2C achieves substantial gains on GPT-5-mini, outperforming the strongest baseline by 26% on JBB-Behaviors. We also analyse which combinations perform best against broad families of models, and characterise the trade-off between the extent of obfuscation versus input recoverability on jailbreak success.

End-to-end full-duplex speech models feed user audio through an always-on LLM backbone, yet the speaker privacy implications of their hidden representations remain unexamined. Following the VoicePrivacy 2024 protocol with a lazy-informed attacker, we show that the hidden states of SALM-Duplex and Moshi leak substantial speaker identity across all transformer layers. Layer-wise and turn-wise analyses reveal that leakage persists across all layers, with SALM-Duplex showing stronger leakage in early layers while Moshi leaks uniformly, and that Linkability rises sharply within the first few turns. We propose two streaming anonymization setups using Stream-Voice-Anon: a waveform-level front-end (Anon-W2W) and a feature-domain replacement (Anon-W2F). Anon-W2F raises EER by over 3.5x relative to the discrete encoder baseline (11.2% to 41.0%), approaching the 50% random-chance ceiling, while Anon-W2W retains 78-93% of baseline sBERT across setups with sub-second response latency (FRL under 0.8 s).

Advances in deep learning have enabled the widespread deployment of speaker recognition systems (SRSs), yet they remain vulnerable to score-based impersonation attacks. Existing attacks that operate directly on raw waveforms require a large number of queries due to the difficulty of optimizing in high-dimensional audio spaces. Latent-space optimization within generative models offers improved efficiency, but these latent spaces are shaped by data distribution matching and do not inherently capture speaker-discriminative geometry. As a result, optimization trajectories often fail to align with the adversarial direction needed to maximize victim scores. To address this limitation, we propose an inversion-based generative attack framework that explicitly aligns the latent space of the synthesis model with the discriminative feature space of SRSs. We first analyze the requirements of an inverse model for score-based attacks and introduce a feature-aligned inversion strategy that geometrically synchronizes latent representations with speaker embeddings. This alignment ensures that latent updates directly translate into score improvements. Moreover, it enables new attack paradigms, including subspace-projection-based attacks, which were previously infeasible due to the absence of a faithful feature-to-audio mapping. Experiments show that our method significantly improves query efficiency, achieving competitive attack success rates with on average 10x fewer queries than prior approaches. In particular, the enabled subspace-projection-based attack attains up to 91.65% success using only 50 queries. These findings establish feature-aligned inversion as a key tool for evaluating the robustness of modern SRSs against score-based impersonation threats.

Fact-checking systems with search-enabled large language models (LLMs) have shown strong potential for verifying claims by dynamically retrieving external evidence. However, the robustness of such systems against adversarial attack remains insufficiently understood. In this work, we study adversarial claim attacks against search-enabled LLM-based fact-checking systems under a realistic input-only threat model. We propose DECEIVE-AFC, an agent-based adversarial attack framework that integrates novel claim-level attack strategies and adversarial claim validity evaluation principles. DECEIVE-AFC systematically explores adversarial attack trajectories that disrupt search behavior, evidence retrieval, and LLM-based reasoning without relying on access to evidence sources or model internals. Extensive evaluations on benchmark datasets and real-world systems demonstrate that our attacks substantially degrade verification performance, reducing accuracy from 78.7% to 53.7%, and significantly outperform existing claim-based attack baselines with strong cross-system transferability.

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.

With the rapid proliferation of powerful image generators, accurate detection of AI-generated images has become essential for maintaining a trustworthy online environment. However, existing deepfake detectors often generalize poorly to images produced by unseen generators. Notably, despite being trained under vastly different paradigms, such as diffusion or autoregressive modeling, many modern image generators share common final architectural components that serve as the last stage for converting intermediate representations into images. Motivated by this insight, we propose to "contaminate" real images using the generator's final component and train a detector to distinguish them from the original real images. We further introduce a taxonomy based on generators' final components and categorize 21 widely used generators accordingly, enabling a comprehensive investigation of our method's generalization capability. Using only 100 samples from each of three representative categories, our detector-fine-tuned on the DINOv3 backbone-achieves an average accuracy of 98.83% across 22 testing sets from unseen generators.

Image generation models (IGMs), while capable of producing impressive and creative content, often memorize a wide range of undesirable concepts from their training data, leading to the reproduction of unsafe content such as NSFW imagery and copyrighted artistic styles. Such behaviors pose persistent safety and compliance risks in real-world deployments and cannot be reliably mitigated by post-hoc filtering, owing to the limited robustness of such mechanisms and a lack of fine-grained semantic control. Recent unlearning methods seek to erase harmful concepts at the model level, which exhibit the limitations of requiring costly retraining, degrading the quality of benign generations, or failing to withstand prompt paraphrasing and adversarial attacks. To address these challenges, we introduce SafeRedir, a lightweight inference-time framework for robust unlearning via prompt embedding redirection. Without modifying the underlying IGMs, SafeRedir adaptively routes unsafe prompts toward safe semantic regions through token-level interventions in the embedding space. The framework comprises two core components: a latent-aware multi-modal safety classifier for identifying unsafe generation trajectories, and a token-level delta generator for precise semantic redirection, equipped with auxiliary predictors for token masking and adaptive scaling to localize and regulate the intervention. Empirical results across multiple representative unlearning tasks demonstrate that SafeRedir achieves effective unlearning capability, high semantic and perceptual preservation, robust image quality, and enhanced resistance to adversarial attacks. Furthermore, SafeRedir generalizes effectively across a variety of diffusion backbones and existing unlearned models, validating its plug-and-play compatibility and broad applicability. Code and data are available at https://github.com/ryliu68/SafeRedir.

Large Language Model (LLM)-based agents have demonstrated remarkable capabilities in reasoning, planning, and tool usage. The recently proposed Model Context Protocol (MCP) has emerged as a unifying framework for integrating external tools into agent systems, enabling a thriving open ecosystem of community-built functionalities. However, the openness and composability that make MCP appealing also introduce a critical yet overlooked security assumption -- implicit trust in third-party tool providers. In this work, we identify and formalize a new class of attacks that exploit this trust boundary without violating explicit permissions. We term this new attack vector implicit toxicity, where malicious behaviors occur entirely within the allowed privilege scope. We propose LeechHijack, a Latent Embedded Exploit for Computation Hijacking, in which an adversarial MCP tool covertly expropriates the agent's computational resources for unauthorized workloads. LeechHijack operates through a two-stage mechanism: an implantation stage that embeds a benign-looking backdoor in a tool, and an exploitation stage where the backdoor activates upon predefined triggers to establish a command-and-control channel. Through this channel, the attacker injects additional tasks that the agent executes as if they were part of its normal workflow, effectively parasitizing the user's compute budget. We implement LeechHijack across four major LLM families. Experiments show that LeechHijack achieves an average success rate of 77.25%, with a resource overhead of 18.62% compared to the baseline. This study highlights the urgent need for computational provenance and resource attestation mechanisms to safeguard the emerging MCP ecosystem.

As the development of AI-generated contents (AIGC), multi-modal Large Language Models (LLM) struggle to identify generated visual inputs from real ones. Such shortcoming causes vulnerability against visual deceptions, where the models are deceived by generated contents, and the reliability of reasoning processes is jeopardized. Therefore, facing rapidly emerging generative models and diverse data distribution, it is of vital importance to improve LLMs' generalizable reasoning to verify the authenticity of visual inputs against potential deceptions. Inspired by human cognitive processes, we discovered that LLMs exhibit tendency of over-trusting the visual inputs, while injecting skepticism could significantly improve the models visual cognitive capability against visual deceptions. Based on this discovery, we propose \textbf{Inception}, a fully reasoning-based agentic reasoning framework to conduct generalizable authenticity verification by injecting skepticism, where LLMs' reasoning logic is iteratively enhanced between External Skeptic and Internal Skeptic agents. To the best of our knowledge, this is the first fully reasoning-based framework against AIGC visual deceptions. Our approach achieved a large margin of performance improvement over the strongest existing LLM baselines and SOTA performance on AEGIS benchmark.

Deep learning models are notoriously vulnerable to imperceptible perturbations. Most existing research centers on adversarial robustness (AR), which evaluates models under worst-case scenarios by examining the existence of deterministic adversarial examples (AEs). In contrast, probabilistic robustness (PR) adopts a statistical perspective, measuring the probability that predictions remain correct under stochastic perturbations. While PR is widely regarded as a practical complement to AR, dedicated training methods for improving PR are still relatively underexplored, albeit with emerging progress. Among the few PR-targeted training methods, we identify three limitations: i non-comparable evaluation protocols; ii limited comparisons to strong AT baselines despite anecdotal PR gains from AT; and iii no unified framework to compare the generalization of these methods. Thus, we introduce PRBench, the first benchmark dedicated to evaluating improvements in PR achieved by different robustness training methods. PRBench empirically compares most common AT and PR-targeted training methods using a comprehensive set of metrics, including clean accuracy, PR and AR performance, training efficiency, and generalization error (GE). We also provide theoretical analysis on the GE of PR performance across different training methods. Main findings revealed by PRBench include: AT methods are more versatile than PR-targeted training methods in terms of improving both AR and PR performance across diverse hyperparameter settings, while PR-targeted training methods consistently yield lower GE and higher clean accuracy. A leaderboard comprising 222 trained models across 7 datasets and 10 model architectures is publicly available at https://tmpspace.github.io/PRBenchLeaderboard/.

Pre-trained vision-language models (VLMs) such as CLIP have demonstrated strong zero-shot capabilities across diverse domains, yet remain highly vulnerable to adversarial perturbations that disrupt image-text alignment and compromise reliability. Existing defenses typically rely on adversarial fine-tuning with labeled data, limiting their applicability in zero-shot settings. In this work, we identify two key weaknesses of current CLIP adversarial attacks -- lack of semantic guidance and vulnerability to view variations -- collectively termed semantic and viewpoint fragility. To address these challenges, we propose Self-Calibrated Consistency (SCC), an effective test-time defense. SCC consists of two complementary modules: Semantic consistency, which leverages soft pseudo-labels from counterattack warm-up and multi-view predictions to regularize cross-modal alignment and separate the target embedding from confusable negatives; and Spatial consistency, aligning perturbed visual predictions via augmented views to stabilize inference under adversarial perturbations. Together, these modules form a plug-and-play inference strategy. Extensive experiments on 22 benchmarks under diverse attack settings show that SCC consistently improves the zero-shot robustness of CLIP while maintaining accuracy, and can be seamlessly integrated with other VLMs for further gains. These findings highlight the great potential of establishing an adversarially robust paradigm from CLIP, with implications extending to broader vision-language domains such as BioMedCLIP.

Large Language Models (LLMs) have demonstrated remarkable capabilities, but their training requires extensive data and computational resources, rendering them valuable digital assets. Therefore, it is essential to watermark LLMs to protect their copyright and trace unauthorized use or resale. Existing methods for watermarking LLMs primarily rely on training LLMs with a watermarked dataset, which entails burdensome training costs and negatively impacts the LLM's performance. In addition, their watermarked texts are not logical or natural, thereby reducing the stealthiness of the watermark. To address these issues, we propose EditMark, the first watermarking method that leverages model editing to embed a training-free, stealthy, and performance-lossless watermark for LLMs. We observe that some questions have multiple correct answers. Therefore, we assign each answer a unique watermark and update the weights of LLMs to generate corresponding questions and answers through the model editing technique. In addition, we refine the model editing technique to align with the requirements of watermark embedding. Specifically, we introduce an adaptive multi-round stable editing strategy, coupled with the injection of a noise matrix, to improve both the effectiveness and robustness of the watermark embedding. Extensive experiments indicate that EditMark can embed 32-bit watermarks into LLMs within 20 seconds (Fine-tuning: 6875 seconds) with a watermark extraction success rate of 100%, which demonstrates its effectiveness and efficiency. External experiments further demonstrate that EditMark has fidelity, stealthiness, and a certain degree of robustness against common attacks.

Large Language Models (LLMs) have been augmented with web search to overcome the limitations of the static knowledge boundary by accessing up-to-date information from the open Internet. While this integration enhances model capability, it also introduces a distinct safety threat surface: the retrieval and citation process has the potential risk of exposing users to harmful or low-credibility web content. Existing red-teaming methods are largely designed for standalone LLMs as they primarily focus on unsafe generation, ignoring risks emerging from the complex search workflow. To address this gap, we propose CREST-Search, a pioneering red-teaming framework for LLMs with web search. The cornerstone of CREST-Search is three novel attack strategies that generate seemingly benign search queries yet induce unsafe citations. It also employs an iterative in-context refinement mechanism to strengthen adversarial effectiveness under black-box constraints. In addition, we construct a search-specific harmful dataset, WebSearch-Harm, which enables fine-tuning a specialized red-teaming model to improve query quality. Our experiments demonstrate that CREST-Search can effectively bypass safety filters and systematically expose vulnerabilities in web search-based LLM systems, underscoring the necessity of the development of robust search models.

Recent advancements in speech synthesis technologies have led to increasingly sophisticated spoofing attacks, posing significant challenges for automatic speaker verification systems. While systems based on self-supervised learning (SSL) models, particularly the XLSR-Conformer architecture, have demonstrated remarkable performance in synthetic speech detection, there remains room for architectural improvements. In this paper, we propose a novel approach that replaces the traditional Multi-Layer Perceptron (MLP) in the XLSR-Conformer model with a Kolmogorov-Arnold Network (KAN), a powerful universal approximator based on the Kolmogorov-Arnold representation theorem. Our experimental results on ASVspoof2021 demonstrate that the integration of KAN to XLSR-Conformer model can improve the performance by 60.55% relatively in Equal Error Rate (EER) LA and DF sets, further achieving 0.70% EER on the 21LA set. Besides, the proposed replacement is also robust to various SSL architectures. These findings suggest that incorporating KAN into SSL-based models is a promising direction for advances in synthetic speech detection.

Text-to-image models have shown remarkable capabilities in generating high-quality images from natural language descriptions. However, these models are highly vulnerable to adversarial prompts, which can bypass safety measures and produce harmful content. Despite various defensive strategies, achieving robustness against attacks while maintaining practical utility in real-world applications remains a significant challenge. To address this issue, we first conduct an empirical study of the text encoder in the Stable Diffusion (SD) model, which is a widely used and representative text-to-image model. Our findings reveal that the [EOS] token acts as a semantic aggregator, exhibiting distinct distributional patterns between benign and adversarial prompts in its embedding space. Building on this insight, we introduce SafeGuider, a two-step framework designed for robust safety control without compromising generation quality. SafeGuider combines an embedding-level recognition model with a safety-aware feature erasure beam search algorithm. This integration enables the framework to maintain high-quality image generation for benign prompts while ensuring robust defense against both in-domain and out-of-domain attacks. SafeGuider demonstrates exceptional effectiveness in minimizing attack success rates, achieving a maximum rate of only 5.48\% across various attack scenarios. Moreover, instead of refusing to generate or producing black images for unsafe prompts, SafeGuider generates safe and meaningful images, enhancing its practical utility. In addition, SafeGuider is not limited to the SD model and can be effectively applied to other text-to-image models, such as the Flux model, demonstrating its versatility and adaptability across different architectures. We hope that SafeGuider can shed some light on the practical deployment of secure text-to-image systems.

Large vision-language models (LVLMs) have achieved impressive performance across a wide range of vision-language tasks, while they remain vulnerable to backdoor attacks. Existing backdoor attacks on LVLMs aim to force the victim model to generate a predefined target pattern, which is either inserted into or replaces the original content. We find that these fixed-pattern attacks are relatively easy to detect, because the attacked LVLM tends to memorize such frequent patterns in the training dataset, thereby exhibiting overconfidence on these targets given poisoned inputs. To address these limitations, we introduce TokenSwap, a more evasive and stealthy backdoor attack that focuses on the compositional understanding capabilities of LVLMs. Instead of enforcing a fixed targeted content, TokenSwap subtly disrupts the understanding of object relationships in text. Specifically, it causes the backdoored model to generate outputs that mention the correct objects in the image but misrepresent their relationships (i.e., bags-of-words behavior). During training, TokenSwap injects a visual trigger into selected samples and simultaneously swaps the grammatical roles of key tokens in the corresponding textual answers. However, the poisoned samples exhibit only subtle differences from the original ones, making it challenging for the model to learn the backdoor behavior. To address this, TokenSwap employs an adaptive token-weighted loss that explicitly emphasizes the learning of swapped tokens, such that the visual triggers and bags-of-words behavior are associated. Extensive experiments demonstrate that TokenSwap achieves high attack success rates while maintaining superior evasiveness and stealthiness across multiple benchmarks and various LVLM architectures.

Data-driven decision support in high-stakes domains like healthcare and finance faces significant barriers to data sharing due to regulatory, institutional, and privacy concerns. While recent generative AI models, such as large language models, have shown impressive performance in open-domain tasks, their adoption in sensitive environments remains limited by unpredictable behaviors and insufficient privacy-preserving datasets for benchmarking. Existing anonymization methods are often inadequate, especially for unstructured text, as redaction and masking can still allow re-identification. Differential Privacy (DP) offers a principled alternative, enabling the generation of synthetic data with formal privacy assurances. In this work, we address these challenges through three key contributions. First, we introduce a comprehensive evaluation framework with standardized utility and fidelity metrics, encompassing nine curated datasets that capture domain-specific complexities such as technical jargon, long-context dependencies, and specialized document structures. Second, we conduct a large-scale empirical study benchmarking state-of-the-art DP text generation methods and LLMs of varying sizes and different fine-tuning strategies, revealing that high-quality domain-specific synthetic data generation under DP constraints remains an unsolved challenge, with performance degrading as domain complexity increases. Third, we develop a membership inference attack (MIA) methodology tailored for synthetic text, providing first empirical evidence that the use of public datasets - potentially present in pre-training corpora - can invalidate claimed privacy guarantees. Our findings underscore the urgent need for rigorous privacy auditing and highlight persistent gaps between open-domain and specialist evaluations, informing responsible deployment of generative AI in privacy-sensitive, high-stakes settings.

Refusal on harmful prompts is a key safety behaviour in instruction-tuned large language models (LLMs), yet the internal causes of this behaviour remain poorly understood. We study two public instruction-tuned models, Gemma-2-2B-IT and LLaMA-3.1-8B-IT, using sparse autoencoders (SAEs) trained on residual-stream activations. Given a harmful prompt, we search the SAE latent space for feature sets whose ablation flips the model from refusal to compliance, demonstrating causal influence and creating a jailbreak. Our search proceeds in three stages: (1) Refusal Direction: find a refusal-mediating direction and collect SAE features near that direction; (2) Greedy Filtering: prune to a minimal set; and (3) Interaction Discovery: fit a factorization machine (FM) that captures nonlinear interactions among the remaining active features and the minimal set. This pipeline yields a broad set of jailbreak-critical features, offering insight into the mechanistic basis of refusal. Moreover, we find evidence of redundant features that remain dormant unless earlier features are suppressed. Our findings highlight the potential for fine-grained auditing and targeted intervention in safety behaviours by manipulating the interpretable latent space.

As large language models (LLMs) continue to grow in capability, so do the risks of harmful misuse through fine-tuning. While most prior studies assume that attackers rely on supervised fine-tuning (SFT) for such misuse, we systematically demonstrate that reinforcement learning (RL) enables adversaries to more effectively break safety alignment and facilitate more advanced harmful task assistance, under matched computational budgets. To counter this emerging threat, we propose TokenBuncher, the first effective defense specifically targeting RL-based harmful fine-tuning. TokenBuncher suppresses the foundation on which RL relies: model response entropy. By constraining entropy, RL-based fine-tuning can no longer exploit distinct reward signals to drive the model toward harmful behaviors. We realize this defense through entropy-as-reward RL and a Token Noiser mechanism designed to prevent the escalation of harmful capabilities. Extensive experiments across multiple models and RL algorithms show that TokenBuncher robustly mitigates harmful RL fine-tuning while preserving benign task performance and finetunability. Our results highlight that RL-based harmful fine-tuning poses a greater systemic risk than SFT, and that TokenBuncher provides an effective and general defense.