agent-evaluation

Pass

Audited by Gen Agent Trust Hub on Apr 30, 2026

Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill contains examples in SKILL.md using subprocess.run to execute commands (e.g., pytest) on agent-generated code for evaluation purposes.\n- [REMOTE_CODE_EXECUTION]: The instructions in SKILL.md include a coding example where the exec() function is used to run arbitrary code strings produced by an agent.\n- [PROMPT_INJECTION]: The skill describes processing untrusted agent data (outcomes, transcripts) which presents an indirect prompt injection surface.\n
  • Ingestion points: Agent outcomes (outcome[\"code\"], outcome[\"report\"]) and action transcripts in SKILL.md.\n
  • Boundary markers: Not explicitly defined in simplified code examples.\n
  • Capability inventory: Execution of shell commands and dynamic Python code via exec and subprocess in SKILL.md.\n
  • Sanitization: The skill explicitly recommends Docker-based environment isolation for running agent code in SKILL.md.\n- [SAFE]: References to external benchmarks and research (e.g., Anthropic, SWE-bench, WebArena) point to trusted and well-known official sources.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 30, 2026, 05:52 AM
Security Audit — agent-trust-hub — agent-evaluation