generate-appworld-code

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCREDENTIALS_UNSAFEDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill generates Python code based on untrusted inputs which could be manipulated to perform unintended actions.
  • Ingestion points: The skill accepts an 'instruction' and 'bullets' field containing external, potentially untrusted content.
  • Boundary markers: The skill instructions do not specify any delimiters or warnings to isolate user data from the code generation logic.
  • Capability inventory: The generated code has access to sensitive APIs for Gmail (sending/fetching emails), Venmo (payments), and Spotify.
  • Sanitization: There is no requirement for input validation or sanitization before the parameters are used for code generation.
  • [CREDENTIALS_UNSAFE]: The skill documentation and examples promote the use of hardcoded credentials.
  • Evidence: API patterns and examples in SKILL.md use 'username="user@example.com"' and 'password="password"'.
  • Risk: The instructions to 'generate actual implementations' using these patterns instead of environment variables or secure storage promotes insecure development practices.
  • [DYNAMIC_EXECUTION]: The skill is specifically designed to produce raw scripts for automated runtime execution.
  • Evidence: The skill requires raw Python code as output and commands the agent to include lifecycle management calls like 'apis.supervisor.complete_task()', indicating a direct execution pipeline.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 02:41 PM