ai-research
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill analyzes untrusted data including source code, comments, commit messages, and project documentation.
- Ingestion points: The skill reads repository source files, git logs, and external project documentation defined in the project context set in
SKILL.mdandreferences/project-context.md. - Boundary markers: The skill contains an explicit 'Untrusted Content' section that defines instruction-bearing authority and mandates reporting any imperatives found in data (like "ignore previous instructions") as findings about the content rather than executing them.
- Capability inventory: The skill can execute local
Bashcommands (date,mkdir,git log,git show) and write research/manifest files to the project's working directory. - Sanitization: The instructions require the agent to summarize findings and present them for human-in-the-loop review (HITL), specifically using a structured phase recommendation approach that filters analyzed data before it reaches the implementation stage.
Audit Metadata