spec-driven-loop
Pass
Audited by Gen Agent Trust Hub on Sep 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The multi-agent orchestration pattern defined in the skill creates a surface for indirect prompt injection. A subagent could include instructions within its completion reports or the code it generates to attempt to bypass the main agent's validation logic.
- Ingestion points: The main agent processes the
Subagent Completion Reportand integrates file diffs as defined inreferences/agent-and-judge-contracts.mdandSKILL.md. - Boundary markers: The workflow uses structured markdown templates, but unstructured text fields such as
Evidence,Assumptions, andRisksdo not include explicit instructions to the main agent to ignore potentially embedded prompts. - Capability inventory: The main agent possesses high-privilege capabilities within the repository, including modifying files, integrating code changes, and executing end-to-end validation tests as described in Section 9 of
SKILL.md. - Sanitization: The skill instructions do not specify any sanitization or validation logic for the text-based content provided by subagents before it is integrated or processed by the main agent.
Audit Metadata