ICLShield: Exploring and Mitigating In-Context Learning Backdoor Attacks

In-context learning (ICL) has demonstrated remarkable success in large language models (LLMs) due to its adaptability and parameter-free nature. However, it also introduces a critical vulnerability to backdoor attacks, where adversaries can manipulate LLM behaviors by simply poisoning a few ICL demonstrations. In this paper, we propose, for the first time, the dual-learning hypothesis, which posits that LLMs simultaneously learn both the task-relevant latent concepts and backdoor latent concepts within poisoned demonstrations, jointly influencing the probability of model outputs. Through theoretical analysis, we derive an upper bound for ICL backdoor effects, revealing that the vulnerability is dominated by the concept preference ratio between the task and the backdoor. Motivated by these findings, we propose ICLShield, a defense mechanism that dynamically adjusts the concept preference ratio. Our method encourages LLMs to select clean demonstrations during the ICL phase by leveraging confidence and similarity scores, effectively mitigating susceptibility to backdoor attacks. Extensive experiments across multiple LLMs and tasks demonstrate that our method achieves state-of-the-art defense effectiveness, significantly outperforming existing approaches (+26.02% on average). Furthermore, our method exhibits exceptional adaptability and defensive performance even for closed-source models (e.g., GPT-4).

SafeMobile: Chain-level Jailbreak Detection and Automated Evaluation for Multimodal Mobile Agents

With the wide application of multimodal foundation models in intelligent agent systems, scenarios such as mobile device control, intelligent assistant interaction, and multimodal task execution are gradually relying on such large model-driven agents. However, the related systems are also increasingly exposed to potential jailbreak risks. Attackers may induce the agents to bypass the original behavioral constraints through specific inputs, and then trigger certain risky and sensitive operations, such as modifying settings, executing unauthorized commands, or impersonating user identities, which brings new challenges to system security. Existing security measures for intelligent agents still have limitations when facing complex interactions, especially in detecting potentially risky behaviors across multiple rounds of conversations or sequences of tasks. In addition, an efficient and consistent automated methodology to assist in assessing and determining the impact of such risks is currently lacking. This work explores the security issues surrounding mobile multimodal agents, attempts to construct a risk discrimination mechanism by incorporating behavioral sequence information, and designs an automated assisted assessment scheme based on a large language model. Through preliminary validation in several representative high-risk tasks, the results show that the method can improve the recognition of risky behaviors to some extent and assist in reducing the probability of agents being jailbroken. We hope that this study can provide some valuable references for the security risk modeling and protection of multimodal intelligent agent systems.

AGENTSAFE: Benchmarking the Safety of Embodied Agents on Hazardous Instructions

The rapid advancement of vision-language models (VLMs) and their integration into embodied agents have unlocked powerful capabilities for decision-making. However, as these systems are increasingly deployed in real-world environments, they face mounting safety concerns, particularly when responding to hazardous instructions. In this work, we propose AGENTSAFE, the first comprehensive benchmark for evaluating the safety of embodied VLM agents under hazardous instructions. AGENTSAFE simulates realistic agent-environment interactions within a simulation sandbox and incorporates a novel adapter module that bridges the gap between high-level VLM outputs and low-level embodied controls. Specifically, it maps recognized visual entities to manipulable objects and translates abstract planning into executable atomic actions in the environment. Building on this, we construct a risk-aware instruction dataset inspired by Asimovs Three Laws of Robotics, including base risky instructions and mutated jailbroken instructions. The benchmark includes 45 adversarial scenarios, 1,350 hazardous tasks, and 8,100 hazardous instructions, enabling systematic testing under adversarial conditions ranging from perception, planning, and action execution stages.

Pushing the Limits of Safety: A Technical Report on the ATLAS Challenge 2025

Multimodal Large Language Models (MLLMs) have enabled transformative advancements across diverse applications but remain susceptible to safety threats, especially jailbreak attacks that induce harmful outputs. To systematically evaluate and improve their safety, we organized the Adversarial Testing & Large-model Alignment Safety Grand Challenge (ATLAS) 2025}. This technical report presents findings from the competition, which involved 86 teams testing MLLM vulnerabilities via adversarial image-text attacks in two phases: white-box and black-box evaluations. The competition results highlight ongoing challenges in securing MLLMs and provide valuable guidance for developing stronger defense mechanisms. The challenge establishes new benchmarks for MLLM safety evaluation and lays groundwork for advancing safer multimodal AI systems. The code and data for this challenge are openly available at https://github.com/NY1024/ATLAS_Challenge_2025.

ME: Trigger Element Combination Backdoor Attack on Copyright Infringement

The capability of generative diffusion models (DMs) like Stable Diffusion (SD) in replicating training data could be taken advantage of by attackers to launch the Copyright Infringement Attack, with duplicated poisoned image-text pairs. SilentBadDiffusion (SBD) is a method proposed recently, which shew outstanding performance in attacking SD in text-to-image tasks. However, the feasible data resources in this area are still limited, some of them are even constrained or prohibited due to the issues like copyright ownership or inappropriate contents; And not all of the images in current datasets are suitable for the proposed attacking methods; Besides, the state-of-the-art (SoTA) performance of SBD is far from ideal when few generated poisoning samples could be adopted for attacks. In this paper, we raised new datasets accessible for researching in attacks like SBD, and proposed Multi-Element (ME) attack method based on SBD by increasing the number of poisonous visual-text elements per poisoned sample to enhance the ability of attacking, while importing Discrete Cosine Transform (DCT) for the poisoned samples to maintain the stealthiness. The Copyright Infringement Rate (CIR) / First Attack Epoch (FAE) we got on the two new datasets were 16.78% / 39.50 and 51.20% / 23.60, respectively close to or even outperformed benchmark Pokemon and Mijourney datasets. In condition of low subsampling ratio (5%, 6 poisoned samples), MESI and DCT earned CIR / FAE of 0.23% / 84.00 and 12.73% / 65.50, both better than original SBD, which failed to attack at all.

SRD: Reinforcement-Learned Semantic Perturbation for Backdoor Defense in VLMs

Vision-Language Models (VLMs) have achieved remarkable performance in image captioning, but recent studies show they are vulnerable to backdoor attacks. Attackers can inject imperceptible perturbations-such as local pixel triggers or global semantic phrases-into the training data, causing the model to generate malicious, attacker-controlled captions for specific inputs. These attacks are hard to detect and defend due to their stealthiness and cross-modal nature. By analyzing attack samples, we identify two key vulnerabilities: (1) abnormal attention concentration on specific image regions, and (2) semantic drift and incoherence in generated captions. To counter this, we propose Semantic Reward Defense (SRD), a reinforcement learning framework that mitigates backdoor behavior without prior knowledge of triggers. SRD uses a Deep Q-Network to learn policies for applying discrete perturbations (e.g., occlusion, color masking) to sensitive image regions, aiming to disrupt the activation of malicious pathways. We design a semantic fidelity score as the reward signal, which jointly evaluates semantic consistency and linguistic fluency of the output, guiding the agent toward generating robust yet faithful captions. Experiments across mainstream VLMs and datasets show SRD reduces attack success rates to 5.6%, while preserving caption quality on clean inputs with less than 10% performance drop. SRD offers a trigger-agnostic, interpretable defense paradigm against stealthy backdoor threats in multimodal generative models.

T2VShield: Model-Agnostic Jailbreak Defense for Text-to-Video Models

The rapid development of generative artificial intelligence has made text to video models essential for building future multimodal world simulators. However, these models remain vulnerable to jailbreak attacks, where specially crafted prompts bypass safety mechanisms and lead to the generation of harmful or unsafe content. Such vulnerabilities undermine the reliability and security of simulation based applications. In this paper, we propose T2VShield, a comprehensive and model agnostic defense framework designed to protect text to video models from jailbreak threats. Our method systematically analyzes the input, model, and output stages to identify the limitations of existing defenses, including semantic ambiguities in prompts, difficulties in detecting malicious content in dynamic video outputs, and inflexible model centric mitigation strategies. T2VShield introduces a prompt rewriting mechanism based on reasoning and multimodal retrieval to sanitize malicious inputs, along with a multi scope detection module that captures local and global inconsistencies across time and modalities. The framework does not require access to internal model parameters and works with both open and closed source systems. Extensive experiments on five platforms show that T2VShield can reduce jailbreak success rates by up to 35 percent compared to strong baselines. We further develop a human centered audiovisual evaluation protocol to assess perceptual safety, emphasizing the importance of visual level defense in enhancing the trustworthiness of next generation multimodal simulators.

Manipulating Multimodal Agents via Cross-Modal Prompt Injection

The emergence of multimodal large language models has redefined the agent paradigm by integrating language and vision modalities with external data sources, enabling agents to better interpret human instructions and execute increasingly complex tasks. However, in this work, we identify a critical yet previously overlooked security vulnerability in multimodal agents: cross-modal prompt injection attacks. To exploit this vulnerability, we propose CrossInject, a novel attack framework in which attackers embed adversarial perturbations across multiple modalities to align with target malicious content, allowing external instructions to hijack the agent's decision-making process and execute unauthorized tasks. Our approach consists of two key components. First, we introduce Visual Latent Alignment, where we optimize adversarial features to the malicious instructions in the visual embedding space based on a text-to-image generative model, ensuring that adversarial images subtly encode cues for malicious task execution. Subsequently, we present Textual Guidance Enhancement, where a large language model is leveraged to infer the black-box defensive system prompt through adversarial meta prompting and generate an malicious textual command that steers the agent's output toward better compliance with attackers' requests. Extensive experiments demonstrate that our method outperforms existing injection attacks, achieving at least a +26.4% increase in attack success rates across diverse tasks. Furthermore, we validate our attack's effectiveness in real-world multimodal autonomous agents, highlighting its potential implications for safety-critical applications.

Towards Understanding the Safety Boundaries of DeepSeek Models: Evaluation and Findings

This study presents the first comprehensive safety evaluation of the DeepSeek models, focusing on evaluating the safety risks associated with their generated content. Our evaluation encompasses DeepSeek's latest generation of large language models, multimodal large language models, and text-to-image models, systematically examining their performance regarding unsafe content generation. Notably, we developed a bilingual (Chinese-English) safety evaluation dataset tailored to Chinese sociocultural contexts, enabling a more thorough evaluation of the safety capabilities of Chinese-developed models. Experimental results indicate that despite their strong general capabilities, DeepSeek models exhibit significant safety vulnerabilities across multiple risk dimensions, including algorithmic discrimination and sexual content. These findings provide crucial insights for understanding and improving the safety of large foundation models. Our code is available at https://github.com/NY1024/DeepSeek-Safety-Eval.

Adversarial Training for Multimodal Large Language Models against Jailbreak Attacks

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.