deliberation
Pass
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONMETADATA_POISONING
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data in the form of a deliberation 'Topic' and 'Jury focus'. This content is passed to sub-agents for synthesis without explicit sanitization or boundary markers.
- Ingestion points: 'Topic' and 'Jury focus' inputs (SKILL.md).
- Boundary markers: Not present for external inputs.
- Capability inventory: Agent and Read tools.
- Sanitization: Not specified.
- [DYNAMIC_EXECUTION]: The skill utilizes the 'Agent' tool to dynamically spawn sub-instances for the Mediator and Jury roles. This pattern creates autonomous reasoning chains based on previously generated output.
- [METADATA_POISONING]: The skill includes a non-existent citation ('arXiv 2605.00914') within its design documentation. This use of fabricated research to support architectural safety claims is deceptive.
Audit Metadata