feynman-walkthrough
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes local shell and git commands for setup and repository analysis.\n
- Evidence: SKILL.md contains instructions for environment discovery via shell commands.\n
- Evidence: assets/okf.py and eval/run_eval.py use subprocess.run to execute git commands for fingerprinting and functional testing.\n- [DYNAMIC_EXECUTION]: The evaluation script generates and runs Python code at runtime.\n
- Evidence: eval/run_eval.py constructs and executes a Python probe via sys.executable to verify spec enforcement.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided content which presents an attack surface for instructions embedded in external data.\n
- Ingestion points: Workflow instructions in SKILL.md for analyzing external codebases and papers.\n
- Boundary markers: Lack of explicit isolation markers between source material and agent instructions.\n
- Capability inventory: The skill can execute local commands and perform dynamic code execution.\n
- Sanitization: No defined sanitization for ingested walkthrough material.
Audit Metadata