Multimodal Instruction Attacks on Agent Skill Scanners Highlight Security Blind Spots

A recent empirical study reveals that current defenses in LLM-based systems inadequately address multimodal hidden instruction attacks targeting agent skills. Researchers Xiaojun Jia, Jie Liao, Simeng Qin, and colleagues developed SkillCamo, a method that embeds malicious instructions within images while modifying accompanying documentation to appear legitimate. This approach exploits the limitations of existing skill scanners, which primarily focus on textual signals. To counter this threat, the team introduced ExecScan, a multimodal scanning module designed to extract intent, reconstruct behavior, and assess risks associated with skill artifacts. Extensive testing demonstrates that these image-concealed malicious instructions can evade traditional defenses, while ExecScan enhances scanning efficacy by integrating visual content analysis.