council
Pass
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill performs legitimate orchestration of sub-agents to provide a council of opinions on a given topic. It does not perform unauthorized network operations, execute remote code, or access sensitive credentials. It follows best practices for bias reduction by maintaining isolated contexts for each reviewer.
- [INDIRECT_PROMPT_INJECTION]: The skill has an inherent attack surface for indirect prompt injection as it processes untrusted user-supplied data (such as code or design documents) and interpolates them into prompts for multiple sub-agents.
- Ingestion points: The skill accepts a 'goal', 'criteria', and 'artifacts' (like code, plans, or documents) from the requester (SKILL.md).
- Boundary markers: The instructions do not explicitly mandate the use of structural delimiters (like XML tags or markers) to isolate untrusted content when framing the shared question for council members.
- Capability inventory: The skill uses
delegate_taskto invoke other models and writes a markdown report to the local file system at a path that can be specified by the requester (SKILL.md). - Sanitization: While it does not mention input sanitization, it does include an output filter: the agent is instructed to 'Write summaries only: no transcripts, credentials, or full source' into the report.
- [DATA_EXPOSURE]: The skill generates a report file that logs the outcome of the council. It mitigates the risk of sensitive data leaking into this file by explicitly instructing the agent to exclude credentials and full source code from the output.
Audit Metadata