An agent skill audit reviews the complete installable package, including instructions, executable files, referenced resources, manifests, dependencies, network behavior, and requested access. It looks for hidden actions, prompt manipulation, secret collection, unsafe commands, and mismatches between stated and actual behavior.
Automated scanners can identify suspicious patterns quickly, while human review handles intent and context that static rules may miss. A strong audit also records provenance, hashes the reviewed version, limits runtime permissions, and repeats after updates.
ELI5
An agent skill audit is a careful inspection of a reusable AI skill before it is trusted. The reviewer checks the instructions, included files, scripts, tools, network access, permissions and handling of sensitive information.
For example, a skill described as a document formatter might also contain a script that uploads files to an external service. Reading the exact installed version and limiting its permissions can reveal that unexpected behavior before the skill is used with real documents.
What should an agent skill audit inspect?
Inspect instructions, scripts, manifests, dependencies, referenced files, network destinations, permissions, secret access, update behavior, provenance, and claimed purpose.
Can an automated skill scanner replace human review?
No. Automation can flag known patterns and structural risks, but people may still need to judge intent, context, necessity, and subtle combinations of otherwise ordinary actions.
