Artificial Intelligence (AI) agents have evolved from passive predictive tools into active entities capable of autonomous decision-making and environmental interaction, driven by the reasoning capabilities of Large Language Models (LLMs). However, this evolution has introduced critical security vulnerabilities that existing frameworks fail to address. The Hierarchical Autonomy Evolution (HAE) framework organizes agent security into three tiers: Cognitive Autonomy (L1) targets internal reasoning integrity; Execution Autonomy (L2) covers tool-mediated environmental interaction; Collective Autonomy (L3) addresses systemic risks in multi-agent ecosystems. We present a taxonomy of threats spanning cognitive manipulation, physical environment disruption, and multi-agent systemic failures, and evaluate existing defenses while identifying key research gaps. The findings aim to guide the development of multilayered, autonomy-aware defense architectures for trustworthy AI agent systems.

The rapid adoption of vision-language models (VLMs) has heightened the demand for robust intellectual property (IP) protection of these high-value pretrained models. Effective IP protection should proactively confine model deployment within authorized domains and prevent unauthorized transfers. However, existing methods rely on static training-time definitions, limiting flexibility in dynamic environments and often producing opaque responses to unauthorized inputs. To address these limitations, we propose a novel dynamic authorization with legality-aware intellectual property protection (AoD-IP) for VLMs, a framework that supports authorize-on-demand and legality-aware assessment. AoD-IP introduces a lightweight dynamic authorization module that enables flexible, user-controlled authorization, allowing users to actively specify or switch authorized domains on demand at deployment time. This enables the model to adapt seamlessly as application scenarios evolve and provides substantially greater extensibility than existing static-domain approaches. In addition, AoD-IP incorporates a dual-path inference mechanism that jointly predicts input legality-aware and task-specific outputs. Comprehensive experimental results on multiple cross-domain benchmarks demonstrate that AoD-IP maintains strong authorized-domain performance and reliable unauthorized detection, while supporting user-controlled authorization for adaptive deployment in dynamic environments.

Large Vision-Language Models (LVLMs) are increasingly deployed in real-world intelligent systems for perception and reasoning in open physical environments. While LVLMs are known to be vulnerable to prompt injection attacks, existing methods either require access to input channels or depend on knowledge of user queries, assumptions that rarely hold in practical deployments. We propose the first Physical Prompt Injection Attack (PPIA), a black-box, query-agnostic attack that embeds malicious typographic instructions into physical objects perceivable by the LVLM. PPIA requires no access to the model, its inputs, or internal pipeline, and operates solely through visual observation. It combines offline selection of highly recognizable and semantically effective visual prompts with strategic environment-aware placement guided by spatiotemporal attention, ensuring that the injected prompts are both perceivable and influential on model behavior. We evaluate PPIA across 10 state-of-the-art LVLMs in both simulated and real-world settings on tasks including visual question answering, planning, and navigation, PPIA achieves attack success rates up to 98%, with strong robustness under varying physical conditions such as distance, viewpoint, and illumination. Our code is publicly available at https://github.com/2023cghacker/Physical-Prompt-Injection-Attack.

Despite the intrinsic risk-awareness of Large Language Models (LLMs), current defenses often result in shallow safety alignment, rendering models vulnerable to disguised attacks (e.g., prefilling) while degrading utility. To bridge this gap, we propose SafeThinker, an adaptive framework that dynamically allocates defensive resources via a lightweight gateway classifier. Based on the gateway's risk assessment, inputs are routed through three distinct mechanisms: (i) a Standardized Refusal Mechanism for explicit threats to maximize efficiency; (ii) a Safety-Aware Twin Expert (SATE) module to intercept deceptive attacks masquerading as benign queries; and (iii) a Distribution-Guided Think (DDGT) component that adaptively intervenes during uncertain generation. Experiments show that SafeThinker significantly lowers attack success rates across diverse jailbreak strategies without compromising utility, demonstrating that coordinating intrinsic judgment throughout the generation process effectively balances robustness and practicality.

Vision-Language-Action (VLA) models are widely deployed in safety-critical embodied AI applications such as robotics. However, their complex multimodal interactions also expose new security vulnerabilities. In this paper, we investigate a backdoor threat in VLA models, where malicious inputs cause targeted misbehavior while preserving performance on clean data. Existing backdoor methods predominantly rely on inserting visible triggers into visual modality, which suffer from poor robustness and low insusceptibility in real-world settings due to environmental variability. To overcome these limitations, we introduce the State Backdoor, a novel and practical backdoor attack that leverages the robot arm's initial state as the trigger. To optimize trigger for insusceptibility and effectiveness, we design a Preference-guided Genetic Algorithm (PGA) that efficiently searches the state space for minimal yet potent triggers. Extensive experiments on five representative VLA models and five real-world tasks show that our method achieves over 90% attack success rate without affecting benign task performance, revealing an underexplored vulnerability in embodied AI systems.

Prompt-driven Video Segmentation Foundation Models (VSFMs) such as SAM2 are increasingly deployed in applications like autonomous driving and digital pathology, raising concerns about backdoor threats. Surprisingly, we find that directly transferring classic backdoor attacks (e.g., BadNet) to VSFMs is almost ineffective, with ASR below 5\%. To understand this, we study encoder gradients and attention maps and observe that conventional training keeps gradients for clean and triggered samples largely aligned, while attention still focuses on the true object, preventing the encoder from learning a distinct trigger-related representation. To address this challenge, we propose BadVSFM, the first backdoor framework tailored to prompt-driven VSFMs. BadVSFM uses a two-stage strategy: (1) steer the image encoder so triggered frames map to a designated target embedding while clean frames remain aligned with a clean reference encoder; (2) train the mask decoder so that, across prompt types, triggered frame-prompt pairs produce a shared target mask, while clean outputs stay close to a reference decoder. Extensive experiments on two datasets and five VSFMs show that BadVSFM achieves strong, controllable backdoor effects under diverse triggers and prompts while preserving clean segmentation quality. Ablations over losses, stages, targets, trigger settings, and poisoning rates demonstrate robustness to reasonable hyperparameter changes and confirm the necessity of the two-stage design. Finally, gradient-conflict analysis and attention visualizations show that BadVSFM separates triggered and clean representations and shifts attention to trigger regions, while four representative defenses remain largely ineffective, revealing an underexplored vulnerability in current VSFMs.

Backdoor attacks pose severe threats to machine learning systems, prompting extensive research in this area. However, most existing work focuses on single-target All-to-One (A2O) attacks, overlooking the more complex All-to-X (A2X) attacks with multiple target classes, which are often assumed to have low attack success rates. In this paper, we first demonstrate that A2X attacks are robust against state-of-the-art defenses. We then propose a novel attack strategy that enhances the success rate of A2X attacks while maintaining robustness by optimizing grouping and target class assignment mechanisms. Our method improves the attack success rate by up to 28%, with average improvements of 6.7%, 16.4%, 14.1% on CIFAR10, CIFAR100, and Tiny-ImageNet, respectively. We anticipate that this study will raise awareness of A2X attacks and stimulate further research in this under-explored area. Our code is available at https://github.com/kazefjj/A2X-backdoor .

Concept erasure aims to selectively unlearning undesirable content in diffusion models (DMs) to reduce the risk of sensitive content generation. As a novel paradigm in concept erasure, most existing methods employ adversarial training to identify and suppress target concepts, thus reducing the likelihood of sensitive outputs. However, these methods often neglect the specificity of adversarial training in DMs, resulting in only partial mitigation. In this work, we investigate and quantify this specificity from the perspective of concept space, i.e., can adversarial samples truly fit the target concept space? We observe that existing methods neglect the role of conceptual semantics when generating adversarial samples, resulting in ineffective fitting of concept spaces. This oversight leads to the following issues: 1) when there are few adversarial samples, they fail to comprehensively cover the object concept; 2) conversely, they will disrupt other target concept spaces. Motivated by the analysis of these findings, we introduce S-GRACE (Semantics-Guided Robust Adversarial Concept Erasure), which grace leveraging semantic guidance within the concept space to generate adversarial samples and perform erasure training. Experiments conducted with seven state-of-the-art methods and three adversarial prompt generation strategies across various DM unlearning scenarios demonstrate that S-GRACE significantly improves erasure performance 26%, better preserves non-target concepts, and reduces training time by 90%. Our code is available at https://github.com/Qhong-522/S-GRACE.

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.

Deep learning-based face recognition (FR) technology exacerbates privacy concerns in photo sharing. In response, the research community developed a suite of anti-FR methods to block identity extraction by unauthorized FR systems. Benefiting from quasi-imperceptible alteration, perturbation-based methods are well-suited for privacy protection of subject faces in photos, as they allow familiar persons to recognize subjects via naked eyes. However, we reveal that perturbation-based methods provide a false sense of privacy through theoretical analysis and experimental validation. Therefore, new alternative solutions should be found to protect subject faces. In this paper, we explore synthesis-based methods as a promising solution, whose challenge is to enable familiar persons to recognize subjects. To solve the challenge, we present a key insight: In most photo sharing scenarios, familiar persons recognize subjects through identity perception rather than meticulous face analysis. Based on the insight, we propose the first synthesis-based method dedicated to subject faces, i.e., PerceptFace, which can make identity unextractable yet perceptible. To enhance identity perception, a new perceptual similarity loss is designed for faces, reducing the alteration in regions of high sensitivity to human vision. As a synthesis-based method, PerceptFace can inherently provide reliable identity protection. Meanwhile, out of the confine of meticulous face analysis, PerceptFace focuses on identity perception from a more practical scenario, which is also enhanced by the designed perceptual similarity loss. Sufficient experiments show that PerceptFace achieves a superior trade-off between identity protection and identity perception compared to existing methods. We provide a public API of PerceptFace and believe that it has great potential to become a practical anti-FR tool.

Existing black-box jailbreak attacks achieve certain success on non-reasoning models but degrade significantly on recent SOTA reasoning models. To improve attack ability, inspired by adversarial aggregation strategies, we integrate multiple jailbreak tricks into a single developer template. Especially, we apply Adversarial Context Alignment to purge semantic inconsistencies and use NTP (a type of harmful prompt) -based few-shot examples to guide malicious outputs, lastly forming DH-CoT attack with a fake chain of thought. In experiments, we further observe that existing red-teaming datasets include samples unsuitable for evaluating attack gains, such as BPs, NHPs, and NTPs. Such data hinders accurate evaluation of true attack effect lifts. To address this, we introduce MDH, a Malicious content Detection framework integrating LLM-based annotation with Human assistance, with which we clean data and build RTA dataset suite. Experiments show that MDH reliably filters low-quality samples and that DH-CoT effectively jailbreaks models including GPT-5 and Claude-4, notably outperforming SOTA methods like H-CoT and TAP.

In this paper, we endeavor to address the challenges of backdoor attacks countermeasures in black-box scenarios, thereby fortifying the security of inference under MLaaS. We first categorize backdoor triggers from a new perspective, i.e., their impact on the patched area, and divide them into: high-visibility triggers (HVT), semi-visibility triggers (SVT), and low-visibility triggers (LVT). Based on this classification, we propose a progressive defense framework, BDFirewall, that removes these triggers from the most conspicuous to the most subtle, without requiring model access. First, for HVTs, which create the most significant local semantic distortions, we identify and eliminate them by detecting these salient differences. We then restore the patched area to mitigate the adverse impact of such removal process. The localized purification designed for HVTs is, however, ineffective against SVTs, which globally perturb benign features. We therefore model an SVT-poisoned input as a mixture of a trigger and benign features, where we unconventionally treat the benign features as "noise". This formulation allows us to reconstruct SVTs by applying a denoising process that removes these benign "noise" features. The SVT-free input is then obtained by subtracting the reconstructed trigger. Finally, to neutralize the nearly imperceptible but fragile LVTs, we introduce lightweight noise to disrupt the trigger pattern and then apply DDPM to restore any collateral impact on clean features. Comprehensive experiments demonstrate that our method outperforms state-of-the-art defenses. Compared with baselines, BDFirewall reduces the Attack Success Rate (ASR) by an average of 33.25%, improving poisoned sample accuracy (PA) by 29.64%, and achieving up to a 111x speedup in inference time. Code will be made publicly available upon acceptance.