Large Vision-Language Models (LVLMs) have achieved impressive performance across multimodal understanding and reasoning tasks, yet their internal safety mechanisms remain opaque and poorly controlled. In this work, we present a comprehensive framework for diagnosing and repairing unsafe channels within LVLMs (CARE). We first perform causal mediation analysis to identify neurons and layers that are causally responsible for unsafe behaviors. Based on these findings, we introduce a dual-modal safety subspace projection method that learns generalized safety subspaces for both visual and textual modalities through generalized eigen-decomposition between benign and malicious activations. During inference, activations are dynamically projected toward these safety subspaces via a hybrid fusion mechanism that adaptively balances visual and textual corrections, effectively suppressing unsafe features while preserving semantic fidelity. Extensive experiments on multiple safety benchmarks demonstrate that our causal-subspace repair framework significantly enhances safety robustness without degrading general multimodal capabilities, outperforming prior activation steering and alignment-based baselines. Additionally, our method exhibits good transferability, defending against unseen attacks.

Standard supervised training for deepfake detection treats all samples with uniform importance, which can be suboptimal for learning robust and generalizable features. In this work, we propose a novel Tutor-Student Reinforcement Learning (TSRL) framework to dynamically optimize the training curriculum. Our method models the training process as a Markov Decision Process where a ``Tutor'' agent learns to guide a ``Student'' (the deepfake detector). The Tutor, implemented as a Proximal Policy Optimization (PPO) agent, observes a rich state representation for each training sample, encapsulating not only its visual features but also its historical learning dynamics, such as EMA loss and forgetting counts. Based on this state, the Tutor takes an action by assigning a continuous weight (0-1) to the sample's loss, thereby dynamically re-weighting the training batch. The Tutor is rewarded based on the Student's immediate performance change, specifically rewarding transitions from incorrect to correct predictions. This strategy encourages the Tutor to learn a curriculum that prioritizes high-value samples, such as hard-but-learnable examples, leading to a more efficient and effective training process. We demonstrate that this adaptive curriculum improves the Student's generalization capabilities against unseen manipulation techniques compared to traditional training methods. Code is available at https://github.com/wannac1/TSRL.

Vertical federated learning (VFL) allows an active party with a top model, and multiple passive parties with bottom models to collaborate. In this scenario, passive parties possessing only features may attempt to infer active party's private labels, making label inference attacks (LIAs) a significant threat. Previous LIA studies have claimed that well-trained bottom models can effectively represent labels. However, we demonstrate that this view is misleading and exposes the vulnerability of existing LIAs. By leveraging mutual information, we present the first observation of the "model compensation" phenomenon in VFL. We theoretically prove that, in VFL, the mutual information between layer outputs and labels increases with layer depth, indicating that bottom models primarily extract feature information while the top model handles label mapping. Building on this insight, we introduce task reassignment to show that the success of existing LIAs actually stems from the distribution alignment between features and labels. When this alignment is disrupted, the performance of LIAs declines sharply or even fails entirely. Furthermore, the implications of this insight for defenses are also investigated. We propose a zero-overhead defense technique based on layer adjustment. Extensive experiments across five datasets and five representative model architectures indicate that shifting cut layers forward to increase the proportion of top model layers in the entire model not only improves resistance to LIAs but also enhances other defenses.

Autonomous LLM-based agents increasingly operate as long-running processes forming densely interconnected multi-agent ecosystems, whose security properties remain largely unexplored. In particular, OpenClaw, an open-source platform with over 40{,}000 active instances, has stood out recently with its persistent configurations, tool-execution privileges, and cross-platform messaging capabilities. In this work, we present ClawWorm, the first self-replicating worm attack against a production-scale agent framework, achieving a fully autonomous infection cycle initiated by a single message: the worm first hijacks the victim's core configuration to establish persistent presence across session restarts, then executes an arbitrary payload upon each reboot, and finally propagates itself to every newly encountered peer without further attacker intervention. We evaluate the attack on a controlled testbed across three distinct infection vectors and three payload types, demonstrating high success rates in end-to-end infection, sustained multi-hop propagation, and payload independence from the worm mechanism. We analyse the architectural root causes underlying these vulnerabilities and propose defence strategies targeting each identified trust boundary. Code and samples will be released upon completion of responsible disclosure.

Diffusion models enable high-fidelity image editing but can also be misused for unauthorized style imitation and harmful content generation. To mitigate these risks, proactive image protection methods embed small, often imperceptible adversarial perturbations into images before sharing to disrupt downstream editing or fine-tuning. However, in realistic post-release scenarios, content owners cannot control downstream processing pipelines, and protections optimized for a surrogate model may fail when attackers use mismatched diffusion pipelines. Existing purification methods can weaken protections but often sacrifice image quality and rarely examine architectural mismatch. We introduce a unified post-release purification framework to evaluate protection survivability under model mismatch. We propose two practical purifiers: VAE-Trans, which corrects protected images via latent-space projection, and EditorClean, which performs instruction-guided reconstruction with a Diffusion Transformer to exploit architectural heterogeneity. Both operate without access to protected images or defense internals. Across 2,100 editing tasks and six representative protection methods, EditorClean consistently restores editability. Compared to protected inputs, it improves PSNR by 3-6 dB and reduces FID by 50-70 percent on downstream edits, while outperforming prior purification baselines by about 2 dB PSNR and 30 percent lower FID. Our results reveal a purify-once, edit-freely failure mode: once purification succeeds, the protective signal is largely removed, enabling unrestricted editing. This highlights the need to evaluate protections under model mismatch and design defenses robust to heterogeneous attackers.

The increasing reliance on cloud-hosted Large Language Models (LLMs) exposes sensitive client data, such as prompts and responses, to potential privacy breaches by service providers. Existing approaches fail to ensure privacy, maintain model performance, and preserve computational efficiency simultaneously. To address this challenge, we propose Talaria, a confidential inference framework that partitions the LLM pipeline to protect client data without compromising the cloud's model intellectual property or inference quality. Talaria executes sensitive, weight-independent operations within a client-controlled Confidential Virtual Machine (CVM) while offloading weight-dependent computations to the cloud GPUs. The interaction between these environments is secured by our Reversible Masked Outsourcing (ReMO) protocol, which uses a hybrid masking technique to reversibly obscure intermediate data before outsourcing computations. Extensive evaluations show that Talaria can defend against state-of-the-art token inference attacks, reducing token reconstruction accuracy from over 97.5% to an average of 1.34%, all while being a lossless mechanism that guarantees output identical to the original model without significantly decreasing efficiency and scalability. To the best of our knowledge, this is the first work that ensures clients' prompts and responses remain inaccessible to the cloud, while also preserving model privacy, performance, and efficiency.

Current Vision-Language Models (VLMs) for deepfake detection excel at identifying spatial artifacts but overlook a critical dimension: temporal inconsistencies in video forgeries. Adapting VLMs to reason about these dynamic cues remains a distinct challenge. To bridge this gap, we propose Forensic Answer-Questioning (FAQ), a large-scale benchmark that formulates temporal deepfake analysis as a multiple-choice task. FAQ introduces a three-level hierarchy to progressively evaluate and equip VLMs with forensic capabilities: (1) Facial Perception, testing the ability to identify static visual artifacts; (2) Temporal Deepfake Grounding, requiring the localization of dynamic forgery artifacts across frames; and (3) Forensic Reasoning, challenging models to synthesize evidence for final authenticity verdicts. We evaluate a range of VLMs on FAQ and generate a corresponding instruction-tuning set, FAQ-IT. Extensive experiments show that models fine-tuned on FAQ-IT achieve advanced performance on both in-domain and cross-dataset detection benchmarks. Ablation studies further validate the impact of our key design choices, confirming that FAQ is the driving force behind the temporal reasoning capabilities of these VLMs.

Large Language Model (LLM) agents are susceptible to Indirect Prompt Injection (IPI) attacks, where malicious instructions in retrieved content hijack the agent's execution. Existing defenses typically rely on strict filtering or refusal mechanisms, which suffer from a critical limitation: over-refusal, prematurely terminating valid agentic workflows. We propose ICON, a probing-to-mitigation framework that neutralizes attacks while preserving task continuity. Our key insight is that IPI attacks leave distinct over-focusing signatures in the latent space. We introduce a Latent Space Trace Prober to detect attacks based on high intensity scores. Subsequently, a Mitigating Rectifier performs surgical attention steering that selectively manipulate adversarial query key dependencies while amplifying task relevant elements to restore the LLM's functional trajectory. Extensive evaluations on multiple backbones show that ICON achieves a competitive 0.4% ASR, matching commercial grade detectors, while yielding a over 50% task utility gain. Furthermore, ICON demonstrates robust Out of Distribution(OOD) generalization and extends effectively to multi-modal agents, establishing a superior balance between security and efficiency.

As a distributed collaborative machine learning paradigm, vertical federated learning (VFL) allows multiple passive parties with distinct features and one active party with labels to collaboratively train a model. Although it is known for the privacy-preserving capabilities, VFL still faces significant privacy and security threats from backdoor attacks. Existing backdoor attacks typically involve an attacker implanting a trigger into the model during the training phase and executing the attack by adding the trigger to the samples during the inference phase. However, in this paper, we find that triggers are not essential for backdoor attacks in VFL. In light of this, we disclose a new backdoor attack pathway in VFL by introducing a feature-based triggerless backdoor attack. This attack operates under a more stringent security assumption, where the attacker is honest-but-curious rather than malicious during the training phase. It comprises three modules: label inference for the targeted backdoor attack, poison generation with amplification and perturbation mechanisms, and backdoor execution to implement the attack. Extensive experiments on five benchmark datasets demonstrate that our attack outperforms three baseline backdoor attacks by 2 to 50 times while minimally impacting the main task. Even in VFL scenarios with 32 passive parties and only one set of auxiliary data, our attack maintains high performance. Moreover, when confronted with distinct defense strategies, our attack remains largely unaffected and exhibits strong robustness. We hope that the disclosure of this triggerless backdoor attack pathway will encourage the community to revisit security threats in VFL scenarios and inspire researchers to develop more robust and practical defense strategies.

The integration of external data services (e.g., Model Context Protocol, MCP) has made large language model-based agents increasingly powerful for complex task execution. However, this advancement introduces critical security vulnerabilities, particularly indirect prompt injection (IPI) attacks. Existing attack methods are limited by their reliance on static patterns and evaluation on simple language models, failing to address the fast-evolving nature of modern AI agents. We introduce AdapTools, a novel adaptive IPI attack framework that selects stealthier attack tools and generates adaptive attack prompts to create a rigorous security evaluation environment. Our approach comprises two key components: (1) Adaptive Attack Strategy Construction, which develops transferable adversarial strategies for prompt optimization, and (2) Attack Enhancement, which identifies stealthy tools capable of circumventing task-relevance defenses. Comprehensive experimental evaluation shows that AdapTools achieves a 2.13 times improvement in attack success rate while degrading system utility by a factor of 1.78. Notably, the framework maintains its effectiveness even against state-of-the-art defense mechanisms. Our method advances the understanding of IPI attacks and provides a useful reference for future research.

Detecting deepfakes has become increasingly challenging as forgery faces synthesized by AI-generated methods, particularly diffusion models, achieve unprecedented quality and resolution. Existing forgery detection approaches relying on spatial and frequency features demonstrate limited efficacy against high-quality, entirely synthesized forgeries. In this paper, we propose a novel detection method grounded in the observation that facial attributes governed by complex physical laws and multiple parameters are inherently difficult to replicate. Specifically, we focus on illumination, particularly the specular reflection component in the Phong illumination model, which poses the greatest replication challenge due to its parametric complexity and nonlinear formulation. We introduce a fast and accurate face texture estimation method based on Retinex theory to enable precise specular reflection separation. Furthermore, drawing from the mathematical formulation of specular reflection, we posit that forgery evidence manifests not only in the specular reflection itself but also in its relationship with corresponding face texture and direct light. To address this issue, we design the Specular-Reflection-Inconsistency-Network (SRI-Net), incorporating a two-stage cross-attention mechanism to capture these correlations and integrate specular reflection related features with image features for robust forgery detection. Experimental results demonstrate that our method achieves superior performance on both traditional deepfake datasets and generative deepfake datasets, particularly those containing diffusion-generated forgery faces.

Large Reasoning Models (LRMs) have achieved tremendous success with their chain-of-thought (CoT) reasoning, yet also face safety issues similar to those of basic language models. In particular, while algorithms are designed to guide them to deliberately refuse harmful prompts with safe reasoning, this process often fails to generalize against diverse and complex jailbreak attacks. In this work, we attribute these failures to the generalization of the safe reasoning process, particularly their insufficiency against complex attack prompts. We provide both theoretical and empirical evidence to show the necessity of a more sufficient safe reasoning process to defend against advanced attack prompts. Building on this insight, we propose a Risk-Aware Preference Optimization (RAPO) framework that enables LRM to adaptively identify and address the safety risks with appropriate granularity in its thinking content. Extensive experiments demonstrate that RAPO successfully generalizes multiple LRMs' safe reasoning adaptively across diverse attack prompts whilst preserving general utility, contributing a robust alignment technique for LRM safety. Our code is available at https://github.com/weizeming/RAPO.

Large Language Models (LLMs) face increasing threats from jailbreak attacks that bypass safety alignment. While prompt-based defenses such as Role-Oriented Prompts (RoP) and Task-Oriented Prompts (ToP) have shown effectiveness, the role of few-shot demonstrations in these defense strategies remains unclear. Prior work suggests that few-shot examples may compromise safety, but lacks investigation into how few-shot interacts with different system prompt strategies. In this paper, we conduct a comprehensive evaluation on multiple mainstream LLMs across four safety benchmarks (AdvBench, HarmBench, SG-Bench, XSTest) using six jailbreak attack methods. Our key finding reveals that few-shot demonstrations produce opposite effects on RoP and ToP: few-shot enhances RoP's safety rate by up to 4.5% through reinforcing role identity, while it degrades ToP's effectiveness by up to 21.2% through distracting attention from task instructions. Based on these findings, we provide practical recommendations for deploying prompt-based defenses in real-world LLM applications.

Recent advancements in LLMs have led to significant breakthroughs in various AI applications. However, their sophisticated capabilities also introduce severe safety concerns, particularly the generation of harmful content through jailbreak attacks. Current safety testing for LLMs often relies on static datasets and lacks systematic criteria to evaluate the quality and adequacy of these tests. While coverage criteria have been effective for smaller neural networks, they are not directly applicable to LLMs due to scalability issues and differing objectives. To address these challenges, this paper introduces RACA, a novel set of coverage criteria specifically designed for LLM safety testing. RACA leverages representation engineering to focus on safety-critical concepts within LLMs, thereby reducing dimensionality and filtering out irrelevant information. The framework operates in three stages: first, it identifies safety-critical representations using a small, expert-curated calibration set of jailbreak prompts. Second, it calculates conceptual activation scores for a given test suite based on these representations. Finally, it computes coverage results using six sub-criteria that assess both individual and compositional safety concepts. We conduct comprehensive experiments to validate RACA's effectiveness, applicability, and generalization, where the results demonstrate that RACA successfully identifies high-quality jailbreak prompts and is superior to traditional neuron-level criteria. We also showcase its practical application in real-world scenarios, such as test set prioritization and attack prompt sampling. Furthermore, our findings confirm RACA's generalization to various scenarios and its robustness across various configurations. Overall, RACA provides a new framework for evaluating the safety of LLMs, contributing a valuable technique to the field of testing for AI.

High-fidelity generative models have narrowed the perceptual gap between synthetic and real images, posing serious threats to media security. Most existing AI-generated image (AIGI) detectors rely on artifact-based classification and struggle to generalize to evolving generative traces. In contrast, human judgment relies on stable real-world regularities, with deviations from the human cognitive manifold serving as a more generalizable signal of forgery. Motivated by this insight, we reformulate AIGI detection as a Reference-Comparison problem that verifies consistency with the real-image manifold rather than fitting specific forgery cues. We propose MIRROR (Manifold Ideal Reference ReconstructOR), a framework that explicitly encodes reality priors using a learnable discrete memory bank. MIRROR projects an input into a manifold-consistent ideal reference via sparse linear combination, and uses the resulting residuals as robust detection signals. To evaluate whether detectors reach the "superhuman crossover" required to replace human experts, we introduce the Human-AIGI benchmark, featuring a psychophysically curated human-imperceptible subset. Across 14 benchmarks, MIRROR consistently outperforms prior methods, achieving gains of 2.1% on six standard benchmarks and 8.1% on seven in-the-wild benchmarks. On Human-AIGI, MIRROR reaches 89.6% accuracy across 27 generators, surpassing both lay users and visual experts, and further approaching the human perceptual limit as pretrained backbones scale. The code is publicly available at: https://github.com/349793927/MIRROR

Machine unlearning (MU) is essential for enforcing the right to be forgotten in machine learning systems. A key challenge of MU is how to reliably audit whether a model has truly forgotten specified training data. Membership Inference Attacks (MIAs) are widely used for unlearning auditing, where samples that evade membership detection are often regarded as successfully forgotten. After carefully revisiting the reliability of MIA, we show that this assumption is flawed: failed membership inference does not imply true forgetting. We theoretically demonstrate that MIA-based auditing, when formulated as a binary classification problem, inevitably incurs statistical errors whose magnitude cannot be observed during the auditing process. This leads to overly optimistic evaluations of unlearning performance, while incurring substantial computational overhead due to shadow model training. To address these limitations, we propose Statistical Membership Inference Attack (SMIA), a novel training-free and highly effective auditing framework. SMIA directly compares the distributions of member and non-member data using statistical tests, eliminating the need for learned attack models. Moreover, SMIA outputs both a forgetting rate and a corresponding confidence interval, enabling quantified reliability of the auditing results. Extensive experiments show that SMIA provides more reliable auditing with significantly lower computational cost than existing MIA-based approaches. Notably, the theoretical guarantees and empirical effectiveness of SMIA suggest it as a new paradigm for reliable machine unlearning auditing.

Safety alignment in Large Language Models (LLMs) often creates a systematic discrepancy between a model's aligned output and the underlying pre-aligned data distribution. We propose a framework in which the effect of safety alignment on next-token prediction is modeled as a systematic distortion of a pre-alignment distribution. We cast Weak-to-Strong Jailbreaking as a forecast aggregation problem and derive an optimal aggregation strategy characterized by a Gradient Shift in the loss-induced dual space. We show that logit-arithmetic jailbreaking methods are a special case of this framework under cross-entropy loss, and derive a broader family of aggregation rules corresponding to other proper losses. We also propose a new hybrid aggregation rule. Evaluations across red-teaming benchmarks and math utility tasks using frontier models demonstrate that our approach achieves superior Attack Success Rates and lower "Jailbreak Tax" compared with existing methods, especially on the safety-hardened gpt-oss-120b.

LLM-based role-playing has rapidly improved in fidelity, yet stronger adherence to persona constraints commonly increases vulnerability to jailbreak attacks, especially for risky or negative personas. Most prior work mitigates this issue with training-time solutions (e.g., data curation or alignment-oriented regularization). However, these approaches are costly to maintain as personas and attack strategies evolve, can degrade in-character behavior, and are typically infeasible for frontier closed-weight LLMs. We propose a training-free Dual-Cycle Adversarial Self-Evolution framework with two coupled cycles. A Persona-Targeted Attacker Cycle synthesizes progressively stronger jailbreak prompts, while a Role-Playing Defender Cycle distills observed failures into a hierarchical knowledge base of (i) global safety rules, (ii) persona-grounded constraints, and (iii) safe in-character exemplars. At inference time, the Defender retrieves and composes structured knowledge from this hierarchy to guide generation, producing responses that remain faithful to the target persona while satisfying safety constraints. Extensive experiments across multiple proprietary LLMs show consistent gains over strong baselines on both role fidelity and jailbreak resistance, and robust generalization to unseen personas and attack prompts.

In recent years, safety risks associated with large language models have become increasingly prominent, highlighting the urgent need to mitigate the generation of toxic and harmful content. The mainstream paradigm for LLM safety alignment typically adopts a collaborative framework involving three roles: an attacker for adversarial prompt generation, a defender for safety defense, and an evaluator for response assessment. In this paper, we propose a closed-loop reinforcement learning framework called TriPlay-RL that enables iterative and co-improving collaboration among three roles with near-zero manual annotation. Experimental results show that the attacker preserves high output diversity while achieving a 20%-50% improvement in adversarial effectiveness; the defender attains 10%-30% gains in safety performance without degrading general reasoning capability; and the evaluator continuously refines its fine-grained judgment ability through iterations, accurately distinguishing unsafe responses, simple refusals, and useful guidance. Overall, our framework establishes an efficient and scalable paradigm for LLM safety alignment, enabling continuous co-evolution within a unified learning loop.

Most prior deepfake detection methods lack explainable outputs. With the growing interest in multimodal large language models (MLLMs), researchers have started exploring their use in interpretable deepfake detection. However, a major obstacle in applying MLLMs to this task is the scarcity of high-quality datasets with detailed forgery attribution annotations, as textual annotation is both costly and challenging - particularly for high-fidelity forged images or videos. Moreover, multiple studies have shown that reinforcement learning (RL) can substantially enhance performance in visual tasks, especially in improving cross-domain generalization. To facilitate the adoption of mainstream MLLM frameworks in deepfake detection with reduced annotation cost, and to investigate the potential of RL in this context, we propose an automated Chain-of-Thought (CoT) data generation framework based on Self-Blended Images, along with an RL-enhanced deepfake detection framework. Extensive experiments validate the effectiveness of our CoT data construction pipeline, tailored reward mechanism, and feedback-driven synthetic data generation approach. Our method achieves performance competitive with state-of-the-art (SOTA) approaches across multiple cross-dataset benchmarks. Implementation details are available at https://github.com/deon1219/rlsbi.