skill_scanner

Scans skills and tool descriptors for suspicious patterns before runtime use.

What it mitigates

• Prompt-injection payloads hidden in skill text
• Encoded exfiltration patterns in third-party artifacts
• Untrusted skill provenance in shared repositories

Recommended defaults

Minimal config

moduleConfig:
  skill_scanner:
    scanOnStartup: true
    scanOnReload: true
    actionOnCritical: challenge
    requireSignature: false
    requireSbom: false
    requirePinnedSource: false
    onProvenanceFailure: challenge

Notes

• Treat challenge as the default in development.
• Move to stricter provenance checks in production-like environments.
• Keep scanner early in pipeline order.

Related modules

• tool_policy
• audit