project-evals

Pass

Audited by Gen Agent Trust Hub on Nov 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The described evaluation workflow involves an AI agent processing a task prompt and modifying application code, which is then analyzed by automated graders. This creates a surface where instructions could potentially be embedded in the agent's output to influence the grading tools.
  • Ingestion points: The targets/<base_app>/task.md file and the agent-modified project files processed by grader.ts (SKILL.md).
  • Boundary markers: The skill encourages the use of neutral hooks and component names, and advises against leaking instructions in prompts to the agent, reducing the risk of accidental obedience to embedded instructions.
  • Capability inventory: The grader.ts scripts utilize Playwright for runtime browser testing and ts-morph for static AST analysis of the agent's code.
  • Sanitization: The guidelines focus on preventing information leakage to the agent ("scrubbing leaks") but do not detail specific sanitization protocols for the agent's output before it is parsed by the grading infrastructure.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 15, 2026, 04:14 AM