architect
Pass
Audited by Gen Agent Trust Hub on Sep 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it processes and synthesizes information from external sources and parallel model outputs.
- Ingestion points: The skill ingests existing source code, grounding artifacts, and design proposals generated by parallel Arena runners as described in SKILL.md Phase A and Phase B.
- Boundary markers: Output is guided by specific templates such as references/rationale-template.md and instructions in references/runner-prompt.md, but it lacks explicit boundary markers or warnings to prevent the agent from following malicious instructions potentially embedded within the processed source code or model-generated proposals.
- Capability inventory: The skill possesses significant capabilities, including the ability to invoke multiple other skills (how, why, arena, interrogate) and the parent agent's ability to implement and write code to the local filesystem during Phase D.
- Sanitization: The skill documentation does not describe any sanitization, validation, or filtering mechanisms for the data ingested from external sources or generated by parallel runners before it is implemented as code.
Audit Metadata