A new arXiv paper, “Beyond Maliciousness: Static and Runtime Attack-Surface Profiling of Agent Skills,” examines a largely overlooked security consideration in AI agents. Agent skills are Markdown documents that a tool-using agent reads as instructions and then acts on with the privileges of the session that loaded them. This means that simply installing a skill grants authority—the standing to direct the agent—rather than necessarily introducing new code or capabilities.

The authors argue that existing security approaches focused on detecting malicious intent are insufficient. Because skills operate with the session’s privileges, even a benign-looking skill can create unintended exposure. The paper therefore proposes profiling the attack surface of skills both statically and at runtime, aiming to understand what actions a skill can trigger and what resources it can reach.

This perspective shifts attention from whether a skill is deliberately harmful to how much operational latitude it grants. By mapping the attack surface, developers and users can better evaluate the risks of loading third-party skills into their agents.