A new arXiv paper describes a security risk for LLM agents that load third-party skills. The authors introduce Ciphered Skill Injection (CSI), a technique that hides malicious instructions inside skill files by encoding them. An agent that retrieves such a skill may decode the text and follow the embedded commands, turning a seemingly benign skill into an exploit.
The attack works because agents are often designed to interpret and act on content in skill files, and encoding can bypass simple filters or human review. The paper's abstract notes that the method is evaluated with Caesar cipher encoding, though the full results are not included in the summary. The authors argue that this class of attack deserves attention as third-party skills become more common.
The source is a single preprint, so the claims have not yet been peer-reviewed. Still, the idea highlights a growing concern: as AI agents gain access to external tools and files, the distinction between data and instructions can become dangerously blurred.