agent-creator
Pass
Audited by Gen Agent Trust Hub on Aug 22, 2026
Risk Level: SAFECOMMAND_EXECUTION
Full Analysis
- [DYNAMIC_EXECUTION]: The template templates/sandbox_runner.py uses compile() and exec() to execute code snippets within a separate namespace. This is an intended architectural pattern for implementing sandboxed execution environments, allowing the agent to run generated code while maintaining process isolation.
- [COMMAND_EXECUTION]: The templates/sandbox_runner.py file utilizes subprocess.Popen to isolate untrusted code execution into a fresh interpreter process. This design allows for the enforcement of wall-clock timeouts and OS-level resource limits (RLIMITs) as documented in the sandboxed execution reference.
- [INDIRECT_PROMPT_INJECTION]: The skill describes an architecture for processing untrusted agent-authored code and tool results. It incorporates security best practices to mitigate injection risks, such as using XML-tagged boundary markers for harness-injected guidance and implementing path-traversal sanitization during file persistence.
Audit Metadata