Security auditor for AI agent skills — catches hidden-Unicode prompt injection, reviewer-subversion, data exfiltration & post-install drift that markdown-only vetters miss.
Three tools with complete input schemas and descriptions, but descriptions are terse (avg 95 chars vs baseline 194). All tools are READ_ONLY or single-write, reducing composition complexity. Naming is clear (scan, baseline, drift) and verb-driven. However, output schemas are not documented in the source, parameter descriptions lack format/constraint details, and error handling guidance is absent. No tool annotations (readOnlyHint/destructiveHint). Schemas are present but minimal, parameters lack examples of valid ranges or formats.
Create a JSON fingerprint baseline of skill files for drift detection
Detect post-install drift by comparing current skill state against a baseline fingerprint
Static security audit of AI agent skills for hidden Unicode, prompt injection, reviewer-subversion, data exfiltration, and code indicators
Output schemas not documented. LLMs cannot plan downstream calls or extract results without knowing what fields scan/baseline/drift return.
Parameter descriptions lack format/constraint details. 'directories' accepts array but no min/max length specified. 'exclude' accepts GLOB but no examples or validation rules provided.
No error handling guidance. Tool descriptions do not explain exit codes (0/1/2) or recovery steps. LLMs cannot self-correct on failure.
Tool annotations missing. No readOnlyHint on scan/drift or destructiveHint on baseline. Agents cannot infer safety/idempotency without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
Descriptions are terse (55 - 60 chars avg vs 194 baseline). 'Create a JSON fingerprint baseline' lacks context on WHEN to call baseline vs drift, or what 'drift' means to an LLM unfamiliar with security auditing.