agent-debate
Pass
Audited by Gen Agent Trust Hub on Sep 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill acts as an orchestrator for multi-agent debates, ingesting external inputs such as alternatives and evidence that could potentially contain malicious instructions.
- Ingestion points: External data enters the agent context through the 'Decision' and 'Shared evidence' sections defined in
references/debate_protocol.mdand processed inSKILL.mdworkflow steps 1 and 3. - Boundary markers: The protocol utilizes structured Markdown templates (Decision frame, Synthesis frame) to delimit untrusted content from system instructions.
- Capability inventory: The skill contains explicit prohibitions against high-risk actions.
SKILL.mdstrictly forbids implementation, source code mutation, commits, merges, or direct writes to canonical memory during the debate process. - Sanitization: The workflow includes mandatory policy checkpoints before spawning reviewers and after each cycle, along with a human-in-the-loop requirement for any final decision or implementation.
Audit Metadata