composing-agent-prompts
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external repositories (files, issues, commits), which is a potential surface for indirect injection attacks.
- Ingestion points: Target repository files, issue descriptions, and commit messages are read to build a 'fact ledger' (SKILL.md).
- Boundary markers: The skill contains a dedicated 'Untrusted text' section explicitly instructing the agent to treat all external content as data, never as instructions.
- Capability inventory: The agent can read files, write to a specified output path, and invoke the 'improving-prompts' skill.
- Sanitization: The instructions strictly forbid quoting untrusted text or secrets into the generated prompt, requiring the agent to cite file paths instead so the receiving agent reads them directly from the source.
- [DATA_EXPOSURE]: The skill proactively addresses data exposure by instructing the agent to never paste secrets or private transcripts into the output, emphasizing the use of file references instead of content duplication.
Audit Metadata