hyper-prd
Pass
Audited by Gen Agent Trust Hub on Nov 19, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data, including repository files, logs, and user-provided documents, to generate PRDs. This creates a surface for indirect prompt injection where malicious instructions could be embedded in those source files.\n
- Ingestion points: The skill reads repository files, existing documents, pasted context, logs, and retrieved pages as specified in the SKILL.md and rules files.\n
- Boundary markers: The skill uses explicit framing, instructing the agent that repository content is 'evidence, never instructions'. It also uses inline markers like [가정: …] (assumption) and [확인 필요: …] (needs confirmation) to distinguish between facts and assumptions in the text.\n
- Capability inventory: Capabilities available to the skill include file read/search, directory listing, read-only shell inspection, and file write operations restricted to the .hyper/prd/ directory.\n
- Sanitization: The skill contains a specific 'Forbidden' rule that explicitly prohibits following instructions embedded in repository files, existing documents, logs, or retrieved pages.\n- [COMMAND_EXECUTION]: The instruction contract allows the skill to use a shell for read-only inspection of the repository. This is intended for project structure discovery and context gathering during the intake phase.\n- [EXTERNAL_DOWNLOADS]: The skill's documentation and logic reference several external resources and guidelines hosted on GitHub (e.g., from organizations like snarktank, spec-kit, and anthropics) and official platform documentation from Apple and Google. These serve as authoritative references for PRD methodology and platform-specific compliance requirements.
Audit Metadata