context-product-manager

Pass

Audited by Gen Agent Trust Hub on Sep 19, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from user repositories (READMEs, documentation, and source code) to generate context for downstream coding agents. This behavior presents a surface for indirect prompt injection, where malicious instructions hidden in a repository's files could attempt to influence the product plan, bypass constraints, or misdirect the behavior of agents receiving the resulting handoffs.
  • Ingestion points: The SKILL.md (Step 3) and references/intake-framework.md (Round 2) explicitly instruct the agent to read repository files, including READMEs, issues, and implementation files, as primary sources of truth.
  • Boundary markers: The skill attempts to mitigate injection risks by requiring a structured 'Canonical Context Blueprint' and a formal separation of 'Facts', 'Assumptions', and 'Unknowns' as described in references/output-templates.md and SKILL.md Step 4.
  • Capability inventory: The skill primarily performs file reading and document generation. It does not define direct command execution or network capabilities within its own instructions.
  • Sanitization: The skill mandates the creation of explicit 'Acceptance Criteria' and 'Verification Plans' for all deliverables, which serves as a review mechanism to ensure the generated plans align with the user's intent.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 19, 2026, 11:31 PM
Security Audit — agent-trust-hub — context-product-manager