openrlhf-training

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.90). The documentation intentionally includes examples that execute untrusted/generated code (via subprocess.run pytest on temp files) and instruct running user-supplied reward/agent Python modules with elevated Docker privileges, which enable remote code execution and system compromise if inputs are not fully trusted.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill's quick-start runs a Docker image pulled at runtime (nvcr.io/nvidia/pytorch:25.02-py3), which fetches and executes remote container code as a required dependency (docker run nvcr.io/nvidia/pytorch:25.02-py3), so it is a runtime external resource that can execute remote code.