consult-claude
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes output from an external AI (Claude Code), creating a potential surface for indirect prompt injection.
- Ingestion points: Claude's responses to questions formulated during the consultation.
- Boundary markers: The skill specifies a rigid format for the final verdict line to help the agent clearly identify the completion of the consultation.
- Capabilities: The agent is authorized to synthesize responses, update project plans (update_plan), and apply suggestions to the codebase.
- Sanitization: The instructions mandate rigorous verification, including cross-referencing suggestions with the repository, checking for prose "drift," and verifying factual claims against their original descriptions before adoption.
- [DATA_EXFILTRATION]: The skill's primary purpose is to share repository context, including file paths, source code, and error messages, with an external analysis tool ($claude-print). This context sharing is the intended core functionality and involves an established external service. The instructions emphasize best practices by advising the agent to default to read-only permissions for this interaction.
Audit Metadata