arena
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill architecture involves ingesting external "Grounding" data and task descriptions which are then processed by multiple parallel subagents. The outputs from these subagents are subsequently read and integrated by a judge agent and the main agent, creating a surface where malicious instructions in the input data or candidate outputs could attempt to influence the agent's behavior.
- Ingestion points: The skill accepts a
TaskandGroundingpaths in Phase B (Fan out) withinSKILL.md. - Boundary markers: The subagents are provided with a structured brief in Phase B to delimit the task, though explicit "ignore instructions" delimiters for the grounding data are not specified.
- Capability inventory: The subagents have file-writing capabilities to specific temporary directories or git worktrees (Phase A). The main agent has capabilities to read these candidate outputs, judge them, and graft content into the base repository (Phases D and E).
- Sanitization: The skill does not perform automated sanitization of external content; instead, it relies on manual verification steps, instructing the agent to "Read every candidate end to end" and "Fold each graft in by hand" to ensure the result remains coherent and safe.
- [EXTERNAL_DOWNLOADS]: The skill metadata references a repository from Cursor (
github.com/cursor/plugins) for attribution purposes.
Audit Metadata