arena
Pass
Audited by Gen Agent Trust Hub on Nov 1, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates a multi-agent workflow where candidate subagents generate outputs that are subsequently read and merged by the main agent. This creates a surface where a subagent could provide malicious instructions that the main agent might inadvertently follow during the 'Graft' or 'Verify' phases.
- Ingestion points: The main agent reads artifacts, rationales, and recommendations from parallel candidate agents in Phase B (Fan out) and Phase C (Cross-judge).
- Boundary markers: The workflow relies on rubric-based scoring but does not define explicit text delimiters or safety instructions to isolate candidate-generated content within the prompt.
- Capability inventory: The skill uses tools for spawning subagents (
subagent spawn), managing git worktrees, and writing task state to~/.local/state/pstack/. - Sanitization: The process includes a manual 'Graft' phase (Phase E) and a 'Verify' phase (Phase F), which serve as filtering steps to prevent the mechanical copying of potentially untrusted or problematic content.
Audit Metadata