model-orchestration
Pass
Audited by Gen Agent Trust Hub on Oct 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and integrate data from sub-agents, creating an attack surface where a sub-agent could potentially influence the primary agent.
- Ingestion points: The primary session reads reports and result files from delegated 'worker' or 'reviewer' tasks as described in the integration logic (SKILL.md, "하위 에이전트의 보고는 주장으로 취급한다").
- Boundary markers: The skill defines a structured 'Contract' for sub-tasks to separate goals, scopes, and verification criteria, serving as a logical delimiter for delegated context.
- Capability inventory: The primary agent has the capability to integrate changes, modify files, and execute verification commands across the workspace based on the information received.
- Sanitization: The skill explicitly mandates that the primary session must treat sub-agent reports as unverified claims and must independently verify actual changes and results rather than relying on sub-agent summaries (SKILL.md, "하위 에이전트의 통과 보고만으로 완료 처리하지 않는다").
Audit Metadata