subagent-driven-development
Pass
Audited by Gen Agent Trust Hub on Jul 10, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill processes implementation plans and task descriptions from external files. Malicious content within these plans could attempt to override subagent behavior. The skill mitigates this by instructing reviewers to verify code independently and not trust reports, though the "Continuous execution" instruction in SKILL.md reduces human oversight.
- Evidence Chain:
- Ingestion points: Implementation plans and task descriptions provided to subagents in implementer-prompt.md, spec-reviewer-prompt.md, and code-quality-reviewer-prompt.md.
- Boundary markers: Present; prompt templates use Markdown headers to separate instructions from untrusted data.
- Capability inventory: Implementation subagents can write code, execute shell tests, and perform git operations.
- Sanitization: No explicit sanitization of plan text is performed before interpolation into prompts.
- [COMMAND_EXECUTION]: The workflow requires subagents to implement code, run tests, and commit work. These actions involve shell command execution. The skill's "Continuous execution" policy in SKILL.md automates a series of these tasks without manual confirmation between steps.
Audit Metadata