Large Language Models (LLMs) have demonstrated impressive performance across various tasks, yet they remain vulnerable to generating toxic content, necessitating detoxification strategies to ensure safe and responsible deployment. Test-time detoxification methods, which typically introduce static or dynamic interventions into LLM representations, offer a promising solution due to their flexibility and minimal invasiveness. However, current approaches often suffer from imprecise interventions, primarily due to their insufficient exploration of the transition space between toxic and non-toxic outputs. To address this challenge, we propose \textsc{A}utoregressive \textsc{R}eward \textsc{G}uided \textsc{R}epresentation \textsc{E}diting (ARGRE), a novel test-time detoxification framework that explicitly models toxicity transitions within the latent representation space, enabling stable and precise reward-guided editing. ARGRE identifies non-toxic semantic directions and interpolates between toxic and non-toxic representations to reveal fine-grained transition trajectories. These trajectories transform sparse toxicity annotations into dense training signals, enabling the construction of an autoregressive reward model that delivers stable and precise editing guidance. At inference, the reward model guides an adaptive two-step editing process to obtain detoxified representations: it first performs directional steering based on expected reward gaps to shift representations toward non-toxic regions, followed by lightweight gradient-based refinements. Extensive experiments across 8 widely used LLMs show that ARGRE significantly outperforms leading baselines in effectiveness (-62.21% toxicity) and efficiency (-47.58% inference time), while preserving the core capabilities of the original model with minimal degradation. Our code is available at the website.

Multimodal Large Language Models (MLLMs) have achieved remarkable success across diverse vision-language tasks, yet their internal decision-making mechanisms remain insufficiently understood. Existing interpretability research has primarily focused on cross-modal attribution, identifying which image regions the model attends to during output generation. However, these approaches often overlook intra-modal dependencies. In the visual modality, attributing importance to isolated image patches ignores spatial context due to limited receptive fields, resulting in fragmented and noisy explanations. In the textual modality, reliance on preceding tokens introduces spurious activations. Failing to effectively mitigate these interference compromises attribution fidelity. To address these limitations, we propose enhancing interpretability by leveraging intra-modal interaction. For the visual branch, we introduce \textit{Multi-Scale Explanation Aggregation} (MSEA), which aggregates attributions over multi-scale inputs to dynamically adjust receptive fields, producing more holistic and spatially coherent visual explanations. For the textual branch, we propose \textit{Activation Ranking Correlation} (ARC), which measures the relevance of contextual tokens to the current token via alignment of their top-$k$ prediction rankings. ARC leverages this relevance to suppress spurious activations from irrelevant contexts while preserving semantically coherent ones. Extensive experiments across state-of-the-art MLLMs and benchmark datasets demonstrate that our approach consistently outperforms existing interpretability methods, yielding more faithful and fine-grained explanations of model behavior.

The safety and reliability of embodied agents rely on accurate and unbiased visual perception. However, existing benchmarks mainly emphasize generalization and robustness under perturbations, while systematic quantification of visual bias remains scarce. This gap limits a deeper understanding of how perception influences decision-making stability. To address this issue, we propose RoboView-Bias, the first benchmark specifically designed to systematically quantify visual bias in robotic manipulation, following a principle of factor isolation. Leveraging a structured variant-generation framework and a perceptual-fairness validation protocol, we create 2,127 task instances that enable robust measurement of biases induced by individual visual factors and their interactions. Using this benchmark, we systematically evaluate three representative embodied agents across two prevailing paradigms and report three key findings: (i) all agents exhibit significant visual biases, with camera viewpoint being the most critical factor; (ii) agents achieve their highest success rates on highly saturated colors, indicating inherited visual preferences from underlying VLMs; and (iii) visual biases show strong, asymmetric coupling, with viewpoint strongly amplifying color-related bias. Finally, we demonstrate that a mitigation strategy based on a semantic grounding layer substantially reduces visual bias by approximately 54.5\% on MOKA. Our results highlight that systematic analysis of visual bias is a prerequisite for developing safe and reliable general-purpose embodied agents.

Visual language modeling for automated driving is emerging as a promising research direction with substantial improvements in multimodal reasoning capabilities. Despite its advanced reasoning abilities, VLM-AD remains vulnerable to serious security threats from adversarial attacks, which involve misleading model decisions through carefully crafted perturbations. Existing attacks have obvious challenges: 1) Physical adversarial attacks primarily target vision modules. They are difficult to directly transfer to VLM-AD systems because they typically attack low-level perceptual components. 2) Adversarial attacks against VLM-AD have largely concentrated on the digital level. To address these challenges, we propose the first Universal Camouflage Attack (UCA) framework for VLM-AD. Unlike previous methods that focus on optimizing the logit layer, UCA operates in the feature space to generate physically realizable camouflage textures that exhibit strong generalization across different user commands and model architectures. Motivated by the observed vulnerability of encoder and projection layers in VLM-AD, UCA introduces a feature divergence loss (FDL) that maximizes the representational discrepancy between clean and adversarial images. In addition, UCA incorporates a multi-scale learning strategy and adjusts the sampling ratio to enhance its adaptability to changes in scale and viewpoint diversity in real-world scenarios, thereby improving training stability. Extensive experiments demonstrate that UCA can induce incorrect driving commands across various VLM-AD models and driving scenarios, significantly surpassing existing state-of-the-art attack methods (improving 30\% in 3-P metrics). Furthermore, UCA exhibits strong attack robustness under diverse viewpoints and dynamic conditions, indicating high potential for practical deployment.

Recently, Large Reasoning Models (LRMs) have demonstrated superior logical capabilities compared to traditional Large Language Models (LLMs), gaining significant attention. Despite their impressive performance, the potential for stronger reasoning abilities to introduce more severe security vulnerabilities remains largely underexplored. Existing jailbreak methods often struggle to balance effectiveness with robustness against adaptive safety mechanisms. In this work, we propose SEAL, a novel jailbreak attack that targets LRMs through an adaptive encryption pipeline designed to override their reasoning processes and evade potential adaptive alignment. Specifically, SEAL introduces a stacked encryption approach that combines multiple ciphers to overwhelm the models reasoning capabilities, effectively bypassing built-in safety mechanisms. To further prevent LRMs from developing countermeasures, we incorporate two dynamic strategies - random and adaptive - that adjust the cipher length, order, and combination. Extensive experiments on real-world reasoning models, including DeepSeek-R1, Claude Sonnet, and OpenAI GPT-o4, validate the effectiveness of our approach. Notably, SEAL achieves an attack success rate of 80.8% on GPT o4-mini, outperforming state-of-the-art baselines by a significant margin of 27.2%. Warning: This paper contains examples of inappropriate, offensive, and harmful content.

Multimodal large language models (MLLMs) have made remarkable strides in cross-modal comprehension and generation tasks. However, they remain vulnerable to jailbreak attacks, where crafted perturbations bypass security guardrails and elicit harmful outputs. In this paper, we present the first adversarial training (AT) paradigm tailored to defend against jailbreak attacks during the MLLM training phase. Extending traditional AT to this domain poses two critical challenges: efficiently tuning massive parameters and ensuring robustness against attacks across multiple modalities. To address these challenges, we introduce Projection Layer Against Adversarial Training (ProEAT), an end-to-end AT framework. ProEAT incorporates a projector-based adversarial training architecture that efficiently handles large-scale parameters while maintaining computational feasibility by focusing adversarial training on a lightweight projector layer instead of the entire model; additionally, we design a dynamic weight adjustment mechanism that optimizes the loss function's weight allocation based on task demands, streamlining the tuning process. To enhance defense performance, we propose a joint optimization strategy across visual and textual modalities, ensuring robust resistance to jailbreak attacks originating from either modality. Extensive experiments conducted on five major jailbreak attack methods across three mainstream MLLMs demonstrate the effectiveness of our approach. ProEAT achieves state-of-the-art defense performance, outperforming existing baselines by an average margin of +34% across text and image modalities, while incurring only a 1% reduction in clean accuracy. Furthermore, evaluations on real-world embodied intelligent systems highlight the practical applicability of our framework, paving the way for the development of more secure and reliable multimodal systems.

Generative large language models are crucial in natural language processing, but they are vulnerable to backdoor attacks, where subtle triggers compromise their behavior. Although backdoor attacks against LLMs are constantly emerging, existing benchmarks remain limited in terms of sufficient coverage of attack, metric system integrity, backdoor attack alignment. And existing pre-trained backdoor attacks are idealized in practice due to resource access constraints. Therefore we establish $\textit{ELBA-Bench}$, a comprehensive and unified framework that allows attackers to inject backdoor through parameter efficient fine-tuning ($\textit{e.g.,}$ LoRA) or without fine-tuning techniques ($\textit{e.g.,}$ In-context-learning). $\textit{ELBA-Bench}$ provides over 1300 experiments encompassing the implementations of 12 attack methods, 18 datasets, and 12 LLMs. Extensive experiments provide new invaluable findings into the strengths and limitations of various attack strategies. For instance, PEFT attack consistently outperform without fine-tuning approaches in classification tasks while showing strong cross-dataset generalization with optimized triggers boosting robustness; Task-relevant backdoor optimization techniques or attack prompts along with clean and adversarial demonstrations can enhance backdoor attack success while preserving model performance on clean samples. Additionally, we introduce a universal toolbox designed for standardized backdoor attack research, with the goal of propelling further progress in this vital area.

Vision-language models (VLMs) have significantly advanced autonomous driving (AD) by enhancing reasoning capabilities; however, these models remain highly susceptible to adversarial attacks. While existing research has explored white-box attacks to some extent, the more practical and challenging black-box scenarios remain largely underexplored due to their inherent difficulty. In this paper, we take the first step toward designing black-box adversarial attacks specifically targeting VLMs in AD. We identify two key challenges for achieving effective black-box attacks in this context: the effectiveness across driving reasoning chains in AD systems and the dynamic nature of driving scenarios. To address this, we propose Cascading Adversarial Disruption (CAD). It first introduces Decision Chain Disruption, which targets low-level reasoning breakdown by generating and injecting deceptive semantics, ensuring the perturbations remain effective across the entire decision-making chain. Building on this, we present Risky Scene Induction, which addresses dynamic adaptation by leveraging a surrogate VLM to understand and construct high-level risky scenarios that are likely to result in critical errors in the current driving contexts. Extensive experiments conducted on multiple AD VLMs and benchmarks demonstrate that CAD achieves state-of-the-art attack effectiveness, significantly outperforming existing methods (+13.43% on average). Moreover, we validate its practical applicability through real-world attacks on AD vehicles powered by VLMs, where the route completion rate drops by 61.11% and the vehicle crashes directly into the obstacle vehicle with adversarial patches. Finally, we release CADA dataset, comprising 18,808 adversarial visual-question-answer pairs, to facilitate further evaluation and research in this critical domain. Our codes and dataset will be available after paper's acceptance.

Website fingerprint (WF) attacks, which covertly monitor user communications to identify the web pages they visit, pose a serious threat to user privacy. Existing WF defenses attempt to reduce the attacker's accuracy by disrupting unique traffic patterns; however, they often suffer from the trade-off between overhead and effectiveness, resulting in less usefulness in practice. To overcome this limitation, we introduce Controllable Website Fingerprint Defense (CWFD), a novel defense perspective based on backdoor learning. CWFD exploits backdoor vulnerabilities in neural networks to directly control the attacker's model by designing trigger patterns based on network traffic. Specifically, CWFD injects only incoming packets on the server side into the target web page's traffic, keeping overhead low while effectively poisoning the attacker's model during training. During inference, the defender can influence the attacker's model through a 'red pill, blue pill' choice: traces with the trigger (red pill) lead to misclassification as the target web page, while normal traces (blue pill) are classified correctly, achieving directed control over the defense outcome. We use the Fast Levenshtein-like distance as the optimization objective to compute trigger patterns that can be effectively associated with our target page. Experiments show that CWFD significantly reduces RF's accuracy from 99% to 6% with 74% data overhead. In comparison, FRONT reduces accuracy to only 97% at similar overhead, while Palette achieves 32% accuracy with 48% more overhead. We further validate the practicality of our method in a real Tor network environment.

The diffusion model has gained significant attention due to its remarkable data generation ability in fields such as image synthesis. However, its strong memorization and replication abilities with respect to the training data also make it a prime target for copyright infringement attacks. This paper provides an in-depth analysis of the spatial similarity of replication in diffusion model and leverages this key characteristic to design a method for detecting poisoning data. By employing a joint assessment of spatial-level and feature-level information from the detected segments, we effectively identify covertly dispersed poisoned samples. Building upon detected poisoning data, we propose a novel defense method specifically targeting copyright infringement attacks by introducing a protection constraint term into the loss function to mitigate the impact of poisoning. Extensive experimental results demonstrate that our approach achieves an average F1 score of 0.709 in detecting copyright infringement backdoors, resulting in an average increase of 68.1% in First-Attack Epoch (FAE) and an average decrease of 51.4% in Copyright Infringement Rate (CIR) of the poisoned model, effectively defending against copyright infringement. Additionally, we introduce the concept of copyright feature inversion, which aids in determining copyright responsibility and expands the application scenarios of defense strategies.