md-review
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted markdown content (drafts and translations) that could contain malicious instructions designed to bypass review criteria or influence the agent's output.
- Ingestion points: Target markdown files and drafts at paths like
docs/src/content/docs/**/*.mdare read directly into the agent context inSKILL.md(Step 1). - Boundary markers: The
reviewer-prompt.mdtemplate uses markdown code fences (triple backticks) to delimit the<DRAFT_CONTENT>and<SOURCE_CONTENT>. This structure is vulnerable if the content itself contains closing backticks, which could allow the content to break out of its block and provide direct instructions to the AI. - Capability inventory: The skill's output JSON (pass/fail) is used as a gate for documentation writeback and publishing, meaning a successful injection could force the approval of non-compliant or malicious content.
- Sanitization: There is no evidence of content escaping, sanitization, or specific boundary protection to prevent code-fence breakouts within the prompt template.
Audit Metadata