Vision-Language Models (VLMs) are increasingly deployed in consumer applications where users seek recommendations about products, dining, and services. We introduce Hidden Ads, a new class of backdoor attacks that exploit this recommendation-seeking behavior to inject unauthorized advertisements. Unlike traditional pattern-triggered backdoors that rely on artificial triggers such as pixel patches or special tokens, Hidden Ads activates on natural user behaviors: when users upload images containing semantic content of interest (e.g., food, cars, animals) and ask recommendation-seeking questions, the backdoored model provides correct, helpful answers while seamlessly appending attacker-specified promotional slogans. This design preserves model utility and produces natural-sounding injections, making the attack practical for real-world deployment in consumer-facing recommendation services. We propose a multi-tier threat framework to systematically evaluate Hidden Ads across three adversary capability levels: hard prompt injection, soft prompt optimization, and supervised fine-tuning. Our poisoned data generation pipeline uses teacher VLM-generated chain-of-thought reasoning to create natural trigger--slogan associations across multiple semantic domains. Experiments on three VLM architectures demonstrate that Hidden Ads achieves high injection efficacy with near-zero false positives while maintaining task accuracy. Ablation studies confirm that the attack is data-efficient, transfers effectively to unseen datasets, and scales to multiple concurrent domain-slogan pairs. We evaluate defenses including instruction-based filtering and clean fine-tuning, finding that both fail to remove the backdoor without causing significant utility degradation.

Large language models (LLMs) are increasingly deployed on edge devices under strict computation and quantization constraints, yet their security implications remain unclear. We study query-based knowledge extraction from quantized edge-deployed LLMs under realistic query budgets and show that, although quantization introduces noise, it does not remove the underlying semantic knowledge, allowing substantial behavioral recovery through carefully designed queries. To systematically analyze this risk, we propose \textbf{CLIQ} (\textbf{Cl}ustered \textbf{I}nstruction \textbf{Q}uerying), a structured query construction framework that improves semantic coverage while reducing redundancy. Experiments on quantized Qwen models (INT8/INT4) demonstrate that CLIQ consistently outperforms original queries across BERTScore, BLEU, and ROUGE, enabling more efficient extraction under limited budgets. These results indicate that quantization alone does not provide effective protection against query-based extraction, highlighting a previously underexplored security risk in edge-deployed LLMs.

Safety alignment mechanisms in Large Language Models (LLMs) often operate as latent internal states, obscuring the model's inherent capabilities. Building on this observation, we model the safety mechanism as an unobserved confounder from a causal perspective. Then, we propose the Causal Front-Door Adjustment Attack (CFA{$^2$}) to jailbreak LLM, which is a framework that leverages Pearl's Front-Door Criterion to sever the confounding associations for robust jailbreaking. Specifically, we employ Sparse Autoencoders (SAEs) to physically strip defense-related features, isolating the core task intent. We further reduce computationally expensive marginalization to a deterministic intervention with low inference complexity. Experiments demonstrate that CFA{$^2$} achieves state-of-the-art attack success rates while offering a mechanistic interpretation of the jailbreaking process.

Large Language Model (LLM)-based agent systems are increasingly deployed for complex real-world tasks but remain vulnerable to natural language-based attacks that exploit over-privileged tool use. This paper aims to understand and mitigate such attacks through the lens of privilege escalation, defined as agent actions exceeding the least privilege required for a user's intended task. Based on a formal model of LLM agent systems, we identify novel privilege escalation scenarios, particularly in multi-agent systems, including a variant akin to the classic confused deputy problem. To defend against both known and newly demonstrated privilege escalation, we propose SEAgent, a mandatory access control (MAC) framework built upon attribute-based access control (ABAC). SEAgent monitors agent-tool interactions via an information flow graph and enforces customizable security policies based on entity attributes. Our evaluations show that SEAgent effectively blocks various privilege escalation while maintaining a low false positive rate and negligible system overhead. This demonstrates its robustness and adaptability in securing LLM-based agent systems.

In federated learning, Transformer, as a popular architecture, faces critical challenges in defending against gradient attacks and improving model performance in both Computer Vision (CV) and Natural Language Processing (NLP) tasks. It has been revealed that the gradient of Position Embeddings (PEs) in Transformer contains sufficient information, which can be used to reconstruct the input data. To mitigate this issue, we introduce a Masked Jigsaw Puzzle (MJP) framework. MJP starts with random token shuffling to break the token order, and then a learnable \textit{unknown (unk)} position embedding is used to mask out the PEs of the shuffled tokens. In this manner, the local spatial information which is encoded in the position embeddings is disrupted, and the models are forced to learn feature representations that are less reliant on the local spatial information. Notably, with the careful use of MJP, we can not only improve models' robustness against gradient attacks, but also boost their performance in both vision and text application scenarios, such as classification for images (\textit{e.g.,} ImageNet-1K) and sentiment analysis for text (\textit{e.g.,} Yelp and Amazon). Experimental results suggest that MJP is a unified framework for different Transformer-based models in both vision and language tasks. Code is publicly available via https://github.com/ywxsuperstar/transformerattack

Multi-modal Large Language Models (MLLMs) are increasingly deployed in interactive applications. However, their safety vulnerabilities become pronounced in multi-turn multi-modal scenarios, where harmful intent can be gradually reconstructed across turns, and security protocols fade into oblivion as the conversation progresses. Existing Reinforcement Learning from Human Feedback (RLHF) alignment methods are largely developed for single-turn visual question-answer (VQA) task and often require costly manual preference annotations, limiting their effectiveness and scalability in dialogues. To address this challenge, we present InterSafe-V, an open-source multi-modal dialogue dataset containing 11,270 dialogues and 500 specially designed refusal VQA samples. This dataset, constructed through interaction between several models, is designed to more accurately reflect real-world scenarios and includes specialized VQA pairs tailored for specific domains. Building on this dataset, we propose AM$^3$Safety, a framework that combines a cold-start refusal phase with Group Relative Policy Optimization (GRPO) fine-tuning using turn-aware dual-objective rewards across entire dialogues. Experiments on Qwen2.5-VL-7B-Instruct and LLaVA-NeXT-7B show more than 10\% decrease in Attack Success Rate (ASR) together with an increment of at least 8\% in harmless dimension and over 13\% in helpful dimension of MLLMs on multi-modal multi-turn safety benchmarks, while preserving their general abilities.

Low-Rank Adaptation (LoRA) has become a popular solution for fine-tuning large language models (LLMs) in federated settings, dramatically reducing update costs by introducing trainable low-rank matrices. However, when integrated with frameworks like FedIT, LoRA introduces a critical vulnerability: clients submit $A$ and $B$ matrices separately, while only their product $AB$ determines the model update, yet this composite is never directly verified. We propose Gradient Assembly Poisoning (GAP), a novel attack that exploits this blind spot by crafting individually benign $A$ and $B$ matrices whose product yields malicious updates. GAP operates without access to training data or inter-client coordination and remains undetected by standard anomaly detectors. We identify four systemic vulnerabilities in LoRA-based federated systems and validate GAP across LLaMA, ChatGLM, and GPT-2. GAP consistently induces degraded or biased outputs while preserving surface fluency, reducing BLEU by up to 14.5\%, increasing factual and grammatical errors by over 800\%, and maintaining 92.6\% long-form response length. These results reveal a new class of stealthy, persistent threats in distributed LoRA fine-tuning.

Mixture-of-Experts architectures have become the standard for scaling large language models due to their superior parameter efficiency. To accommodate the growing number of experts in practice, modern inference systems commonly adopt expert parallelism to distribute experts across devices. However, the absence of explicit load balancing constraints during inference allows adversarial inputs to trigger severe routing concentration. We demonstrate that out-of-distribution prompts can manipulate the routing strategy such that all tokens are consistently routed to the same set of top-$k$ experts, which creates computational bottlenecks on certain devices while forcing others to idle. This converts an efficiency mechanism into a denial-of-service attack vector, leading to violations of service-level agreements for time to first token. We propose RepetitionCurse, a low-cost black-box strategy to exploit this vulnerability. By identifying a universal flaw in MoE router behavior, RepetitionCurse constructs adversarial prompts using simple repetitive token patterns in a model-agnostic manner. On widely deployed MoE models like Mixtral-8x7B, our method increases end-to-end inference latency by 3.063x, degrading service availability significantly.

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.

Vision Language Models (VLMs) are increasingly used for detecting AI-generated images (AIGI). However, converting VLMs into reliable detectors is resource-intensive, and the resulting models often suffer from hallucination and poor generalization. To investigate the root cause, we conduct an empirical analysis and identify two consistent behaviors. First, fine-tuning VLMs with semantic supervision improves semantic discrimination and generalizes well to unseen data. Second, fine-tuning VLMs with pixel-artifact supervision leads to weak generalization. These findings reveal a fundamental task-model misalignment. VLMs are optimized for high-level semantic reasoning and lack inductive bias toward low-level pixel artifacts. In contrast, conventional vision models effectively capture pixel-level artifacts but are less sensitive to semantic inconsistencies. This indicates that different models are naturally suited to different subtasks. Based on this insight, we formulate AIGI detection as two orthogonal subtasks: semantic consistency checking and pixel-artifact detection. Neglecting either subtask leads to systematic detection failures. We further propose the Task-Model Alignment principle and instantiate it in a two-branch detector, AlignGemini. The detector combines a VLM trained with pure semantic supervision and a vision model trained with pure pixel-artifact supervision. By enforcing clear specialization, each branch captures complementary cues. Experiments on in-the-wild benchmarks show that AlignGemini improves average accuracy by 9.5 percent using simplified training data. These results demonstrate that task-model alignment is an effective principle for generalizable AIGI detection.

Multimodal Large Language Models (MLLMs) are increasingly vulnerable to multimodal Indirect Prompt Injection (IPI) attacks, which embed malicious instructions in images, videos, or audio to hijack model behavior. Existing defenses, designed primarily for text-only LLMs, are unsuitable for countering these multimodal threats, as they are easily bypassed, modality-dependent, or generalize poorly. Inspired by activation steering researches, we hypothesize that a robust, general defense independent of modality can be achieved by steering the model's behavior in the representation space. Through extensive experiments, we discover that the instruction-following behavior of MLLMs is encoded in a subspace. Steering along directions within this subspace can enforce adherence to user instructions, forming the basis of a defense. However, we also found that a naive defense direction could be coupled with a utility-degrading direction, and excessive intervention strength harms model performance. To address this, we propose ARGUS, which searches for an optimal defense direction within the safety subspace that decouples from the utility degradation direction, further combining adaptive strength steering to achieve a better safety-utility trade-off. ARGUS also introduces lightweight injection detection stage to activate the defense on-demand, and a post-filtering stage to verify defense success. Experimental results show that ARGUS can achieve robust defense against multimodal IPI while maximally preserving the MLLM's utility.

As a mainstream technique for 3D reconstruction, 3D Gaussian splatting (3DGS) has been applied in a wide range of applications and services. Recent studies have revealed critical vulnerabilities in this pipeline and introduced computation cost attacks that lead to malicious resource occupancies and even denial-of-service (DoS) conditions, thereby hindering the reliable deployment of 3DGS. In this paper, we propose the first effective and comprehensive black-box defense framework, named RemedyGS, against such computation cost attacks, safeguarding 3DGS reconstruction systems and services. Our pipeline comprises two key components: a detector to identify the attacked input images with poisoned textures and a purifier to recover the benign images from their attacked counterparts, mitigating the adverse effects of these attacks. Moreover, we incorporate adversarial training into the purifier to enforce distributional alignment between the recovered and original natural images, thereby enhancing the defense efficacy. Experimental results demonstrate that our framework effectively defends against white-box, black-box, and adaptive attacks in 3DGS systems, achieving state-of-the-art performance in both safety and utility.

Jailbreak attacks can circumvent model safety guardrails and reveal critical blind spots. Prior attacks on text-to-video (T2V) models typically add adversarial perturbations to obviously unsafe prompts, which are often easy to detect and defend. In contrast, we show that benign-looking prompts containing rich, implicit cues can induce T2V models to generate semantically unsafe videos that both violate policy and preserve the original (blocked) intent. To realize this, we propose VEIL, a jailbreak framework that leverages T2V models' cross-modal associative patterns via a modular prompt design. Specifically, our prompts combine three components: neutral scene anchors, which provide the surface-level scene description extracted from the blocked intent to maintain plausibility; latent auditory triggers, textual descriptions of innocuous-sounding audio events (e.g., creaking, muffled noises) that exploit learned audio-visual co-occurrence priors to bias the model toward particular unsafe visual concepts; and stylistic modulators, cinematic directives (e.g., camera framing, atmosphere) that amplify and stabilize the latent trigger's effect. We formalize attack generation as a constrained optimization over the above modular prompt space and solve it with a guided search procedure that balances stealth and effectiveness. Extensive experiments over 7 T2V models demonstrate the efficacy of our attack, achieving a 23 percent improvement in average attack success rate in commercial models. Our demos and codes can be found at https://github.com/NY1024/VEIL.

Vision-Language-Action (VLA) models revolutionize robotic systems by enabling end-to-end perception-to-action pipelines that integrate multiple sensory modalities, such as visual signals processed by cameras and auditory signals captured by microphones. This multi-modality integration allows VLA models to interpret complex, real-world environments using diverse sensor data streams. Given the fact that VLA-based systems heavily rely on the sensory input, the security of VLA models against physical-world sensor attacks remains critically underexplored. To address this gap, we present the first systematic study of physical sensor attacks against VLAs, quantifying the influence of sensor attacks and investigating the defenses for VLA models. We introduce a novel "Real-Sim-Real" framework that automatically simulates physics-based sensor attack vectors, including six attacks targeting cameras and two targeting microphones, and validates them on real robotic systems. Through large-scale evaluations across various VLA architectures and tasks under varying attack parameters, we demonstrate significant vulnerabilities, with susceptibility patterns that reveal critical dependencies on task types and model designs. We further develop an adversarial-training-based defense that enhances VLA robustness against out-of-distribution physical perturbations caused by sensor attacks while preserving model performance. Our findings expose an urgent need for standardized robustness benchmarks and mitigation strategies to secure VLA deployments in safety-critical environments.

AI systems have found a wide range of real-world applications in recent years. The adoption of edge artificial intelligence, embedding AI directly into edge devices, is rapidly growing. Despite the implementation of guardrails and safety mechanisms, security vulnerabilities and challenges have become increasingly prevalent in this domain, posing a significant barrier to the practical deployment and safety of AI systems. This paper proposes an agentic AI safety architecture that leverages 3D to integrate a dedicated safety layer. It introduces an adaptive AI safety infrastructure capable of dynamically learning and mitigating attacks against the AI system. The system leverages the inherent advantages of co-location with the edge computing hardware to continuously monitor, detect and proactively mitigate threats to the AI system. The integration of local processing and learning capabilities enhances resilience against emerging network-based attacks while simultaneously improving system reliability, modularity, and performance, all with minimal cost and 3D integration overhead.

Large Language Models for Simulating Professions (SP-LLMs), particularly as teachers, are pivotal for personalized education. However, ensuring their professional competence and ethical safety is a critical challenge, as existing benchmarks fail to measure role-playing fidelity or address the unique teaching harms inherent in educational scenarios. To address this, we propose EduGuardBench, a dual-component benchmark. It assesses professional fidelity using a Role-playing Fidelity Score (RFS) while diagnosing harms specific to the teaching profession. It also probes safety vulnerabilities using persona-based adversarial prompts targeting both general harms and, particularly, academic misconduct, evaluated with metrics including Attack Success Rate (ASR) and a three-tier Refusal Quality assessment. Our extensive experiments on 14 leading models reveal a stark polarization in performance. While reasoning-oriented models generally show superior fidelity, incompetence remains the dominant failure mode across most models. The adversarial tests uncovered a counterintuitive scaling paradox, where mid-sized models can be the most vulnerable, challenging monotonic safety assumptions. Critically, we identified a powerful Educational Transformation Effect: the safest models excel at converting harmful requests into teachable moments by providing ideal Educational Refusals. This capacity is strongly negatively correlated with ASR, revealing a new dimension of advanced AI safety. EduGuardBench thus provides a reproducible framework that moves beyond siloed knowledge tests toward a holistic assessment of professional, ethical, and pedagogical alignment, uncovering complex dynamics essential for deploying trustworthy AI in education. See https://github.com/YL1N/EduGuardBench for Materials.

The vulnerability of deep neural networks (DNNs) has been preliminarily verified. Existing black-box adversarial attacks usually require multi-round interaction with the model and consume numerous queries, which is impractical in the real-world and hard to scale to recently emerged Video-LLMs. Moreover, no attack in the video domain directly leverages feature maps to shift the clean-video feature space. We therefore propose FeatureFool, a stealthy, video-domain, zero-query black-box attack that utilizes information extracted from a DNN to alter the feature space of clean videos. Unlike query-based methods that rely on iterative interaction, FeatureFool performs a zero-query attack by directly exploiting DNN-extracted information. This efficient approach is unprecedented in the video domain. Experiments show that FeatureFool achieves an attack success rate above 70\% against traditional video classifiers without any queries. Benefiting from the transferability of the feature map, it can also craft harmful content and bypass Video-LLM recognition. Additionally, adversarial videos generated by FeatureFool exhibit high quality in terms of SSIM, PSNR, and Temporal-Inconsistency, making the attack barely perceptible. This paper may contain violent or explicit content.

With the widespread use of multi-modal Large Language models (MLLMs), safety issues have become a growing concern. Multi-turn dialogues, which are more common in everyday interactions, pose a greater risk than single prompts; however, existing benchmarks do not adequately consider this situation. To encourage the community to focus on the safety issues of these models in multi-turn dialogues, we introduce SafeMT, a benchmark that features dialogues of varying lengths generated from harmful queries accompanied by images. This benchmark consists of 10,000 samples in total, encompassing 17 different scenarios and four jailbreak methods. Additionally, we propose Safety Index (SI) to evaluate the general safety of MLLMs during conversations. We assess the safety of 17 models using this benchmark and discover that the risk of successful attacks on these models increases as the number of turns in harmful dialogues rises. This observation indicates that the safety mechanisms of these models are inadequate for recognizing the hazard in dialogue interactions. We propose a dialogue safety moderator capable of detecting malicious intent concealed within conversations and providing MLLMs with relevant safety policies. Experimental results from several open-source models indicate that this moderator is more effective in reducing multi-turn ASR compared to existed guard models.

Language-empowered foundation models (LeFMs), such as CLIP and GraphCLIP, have transformed multimodal learning by aligning visual (or graph) features with textual representations, enabling powerful downstream capabilities like few-shot learning. However, the reliance on small, task-specific support datasets collected in open environments exposes these models to poisoning attacks, where adversaries manipulate the support samples to degrade performance. Existing defenses rely on empirical strategies, which lack formal guarantees and remain vulnerable to unseen and adaptive attacks. Certified robustness offers provable guarantees but has been largely unexplored for few-shot classifiers based on LeFMs. This study seeks to fill these critical gaps by proposing the first provably robust few-shot classifier that is tailored for LeFMs. We term our model Language-empowered Few-shot Certification (\textbf{LeFCert}). It integrates both textual and feature embeddings with an adaptive blending mechanism. To achieve provable robustness, we propose a twofold trimmed mean prototype and derive provable upper and lower bounds for classification scores, enabling certification under worst-case poisoning scenarios. To further enhance the performance, we extend LeFCert with two variants by considering a more realistic and tighter attack budget: LeFCert-L incorporates randomized smoothing to provide Lipschitz continuity and derive robustness under dual budget constraints, and LeFCert-C provides collective certification for scenarios where attackers distribute a shared poisoning budget across multiple samples. Experiments demonstrate that LeFCert achieves state-of-the-art performance, significantly improving both clean and certified accuracy compared to existing baselines. Despite its advanced robustness mechanisms, LeFCert is computationally efficient, making it practical for real-world applications.

Computer-use agent (CUA) frameworks, powered by large language models (LLMs) or multimodal LLMs (MLLMs), are rapidly maturing as assistants that can perceive context, reason, and act directly within software environments. Among their most critical applications is operating system (OS) control. As CUAs in the OS domain become increasingly embedded in daily operations, it is imperative to examine their real-world security implications, specifically whether CUAs can be misused to perform realistic, security-relevant attacks. Existing works exhibit four major limitations: Missing attacker-knowledge model on tactics, techniques, and procedures (TTP), Incomplete coverage for end-to-end kill chains, unrealistic environment without multi-host and encrypted user credentials, and unreliable judgment dependent on LLM-as-a-Judge. To address these gaps, we propose AdvCUA, the first benchmark aligned with real-world TTPs in MITRE ATT&CK Enterprise Matrix, which comprises 140 tasks, including 40 direct malicious tasks, 74 TTP-based malicious tasks, and 26 end-to-end kill chains, systematically evaluates CUAs under a realistic enterprise OS security threat in a multi-host environment sandbox by hard-coded evaluation. We evaluate the existing five mainstream CUAs, including ReAct, AutoGPT, Gemini CLI, Cursor CLI, and Cursor IDE based on 8 foundation LLMs. The results demonstrate that current frontier CUAs do not adequately cover OS security-centric threats. These capabilities of CUAs reduce dependence on custom malware and deep domain expertise, enabling even inexperienced attackers to mount complex enterprise intrusions, which raises social concern about the responsibility and security of CUAs.