Skill safety · 2026-09-23

Why Agent skills need permission disclosure, not just usage instructions

Files, commands, networks, accounts, and external mutations carry different risks that users need to see before execution.

Capability is not permission

“Deploy your app” does not say whether the workflow logs into a cloud account, edits DNS, overwrites configuration, restarts neighboring services, or uploads artifacts. A permission map separates those actions so authorization has a concrete object.

Authorize by outcome

The user authorizes a defined outcome and scope—not unrestricted exploration across every system. High-impact actions require confirmation at execution time and a rollback target.

Verification has boundaries too

Static validation, script testing, live-provider checks, and end-user acceptance are different evidence layers. Public status should state which layer passed and which remains unrun.