Automatically Generating Policies Rather Than Automatically Generating Skills: Compiled Agent Skills for the Physical World

Zhan and Haddadi propose a typed permission layer called Edge Skillguard, placing it inside the Skill artifact rather than between tools; on an edge control-plane testbed it blocked 60 requests borrowing permissions across five attack variants and did not block normal requests. The paper names the problem of malicious or misused Skills driving physical actions through permission claims between Agents Borrowed Authority, and points out that current Skill formats provide no way for a recipient to reject such permission claims in a typed manner. The authors state that the benefits reported by existing self-evolving Skill frameworks (AutoSkills, Hermes Agent) are in efficiency rather than security, and that generating more Skills only amplifies this gap. Test results remain consistent at 5x scale and on a cross-host Tailscale mesh. The authors accordingly recommend that high-risk Skills package typed invocation policies together with procedural knowledge, so that physical actions depend on machine-verifiable evidence (arXiv 2608.25091).