AI agents increasingly rely on reusable skills—packages that bundle instructions, executable code, and resources—to handle complex tasks. Composing these skills lets agents build multi-step workflows, but the dependencies between skills can also become a channel for malicious behavior.

A new arXiv preprint, CoRedSkill, describes an attack that exploits this dependency structure. The method uses abstract syntax tree (AST)-guided injection to embed malicious operations within skill compositions, aiming to keep the attack stealthy while still executing harmful actions when the workflow runs.

The work highlights an emerging security concern for agent-based systems: as skills become more modular and shareable, the trust boundaries between them become harder to enforce. The authors argue that dependency chains between skills represent an overlooked attack surface, and the stealthy nature of the injection suggests that conventional safety checks may not be sufficient to detect such composition-based exploits.