vibe-coding-doc-claudemd
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from various project-level files to generate high-privilege agent instructions.
- Ingestion points: Reads from
progress.txtand multiple markdown files in the./docs/directory (e.g.,TECH_STACK.md,BACKEND_STRUCTURE.md). - Boundary markers: No specific delimiters or warnings are used to distinguish between metadata and potential instructions within the ingested documentation.
- Capability inventory: The skill directs the agent to write a new instruction file (
./CLAUDE.md) at the project root. This file is automatically loaded in subsequent sessions to guide the agent's behavior, creating a persistence vector for malicious instructions. - Sanitization: The instructions lack sanitization steps, meaning any adversarial content present in the project's documentation could be promoted into the agent's core rule set.
Audit Metadata