arena
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill implements a structured workflow for comparing and merging outputs from multiple AI models. It utilizes standard local configuration paths (~/.cursor/rules/pstack-models.mdc) and temporary directories (/tmp) for managing candidate outputs, which are appropriate for developer-focused orchestration tools.
- [INDIRECT_PROMPT_INJECTION]: The skill distributes a user-provided task to multiple sub-agents, creating a processing surface for potential prompt injection.
- Ingestion points: The 'Frame' phase (Phase A) ingests a task prompt that is subsequently passed to all sub-agent runners in the 'Fan out' phase (Phase B).
- Boundary markers: The instructions do not specify explicit delimiters or instruction-override warnings when interpolating the task into sub-agent prompts.
- Capability inventory: The skill spawns background sub-agents (run_in_background: true) and performs file system operations including reading configurations and writing results to temporary workspaces or git worktrees.
- Sanitization: The synthesized result undergoes a dedicated 'Verify' phase (Phase F) to ensure the artifact meets the original rubric and functions correctly, serving as a review checkpoint.
Audit Metadata