verl-e2e-testing

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.90). The required workflow explicitly clones the public verl GitHub repo and instructs fetching public Hugging Face models/checkpoints (MODEL_ID / hf download), downloading the GSM8K dataset via datasets (openai/gsm8k), and pulling a public verlai/verl Docker image — all untrusted third-party content that the agent ingests and executes as part of the run, so it could carry indirect prompt-injection or behavior-influencing content.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly fetches and runs remote code at runtime—e.g., cloning/running the verl repo (https://github.com/verl-project/verl and git clone https://github.com/verl-project/verl.git), pulling the verl Docker image (https://hub.docker.com/r/verlai/verl), and optionally pip-installing from NVIDIA's index (https://pypi.nvidia.com)—which will execute remote code and are required for the run, so they meet the flagging criteria.