blinded-eval
Pass
Audited by Gen Agent Trust Hub on Sep 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill defines a structured methodology for A/B testing changes using isolated temporary directories and subagents to prevent side effects, explicitly warning against accessing data outside the designated workspace.
- [INDIRECT_PROMPT_INJECTION]: The skill defines a framework for processing artifacts generated by external subagents, creating a potential data ingestion surface.
- Ingestion points: The skill reads and evaluates artifacts and files generated by candidate subagents during the parent synthesis phase.
- Boundary markers: The workflow mandates strict isolation within unique working directories or git worktrees and utilizes randomized labels to separate candidate context from the judge.
- Capability inventory: The skill involves spawning isolated subagents, creating local directories, and performing file read operations on candidate outputs.
- Sanitization: Detailed sanitization rules are provided to remove keywords and metadata from directory names and prompts to maintain blinding and prevent model bias.
Audit Metadata