subagent-driven-development
Pass
Audited by Gen Agent Trust Hub on Jun 21, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: No malicious patterns or security vulnerabilities were detected. The skill follows best practices for agentic task delegation and isolation.
- [PROMPT_INJECTION]: The skill manages the attack surface for indirect prompt injection via external development plans through structural mitigations.
- Ingestion points: Implementation plans and task descriptions are read from external files referenced in the workflow (SKILL.md).
- Boundary markers: Task instructions are separated from system instructions using structured markdown headers and context-laying sections in implementer-prompt.md.
- Capability inventory: Sub-agents have permissions to write code, execute tests, and perform git operations within isolated worktrees (SKILL.md, implementer-prompt.md).
- Sanitization: The workflow enforces a verification cycle (spec-reviewer-prompt.md) that explicitly mandates independent code review, instructing the reviewer to 'not trust the report' and 'verify through reading the code', which serves as a mitigation against malicious or incorrect code generation stemming from data poisoning.
- [COMMAND_EXECUTION]: The skill enables autonomous code and test execution as part of its primary development purpose. These operations are conducted within specified directory constraints and are subject to the review cycles mentioned above.
Audit Metadata