SkillCFI applies control-flow integrity to agent skills: it blocked all 240 rule-breaking steps that permission allowlists let through, and none of 53 public skills was enforceable as written
- Takeaway: allowlists stop unauthorized tools, not authorized tools used in the wrong order (e.g. overwrite before validate). Put ordering checks in the harness, not in the prompt.
Sources: arXiv 2611.24185: Control-Flow Integrity for Agent Skills (SkillCFI), Ruan et al. (Nov 18, 2026)