eval-model

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

• Third-party content exposure detected (high risk: 0.85). Outsider-authored free text from public Hugging Face datasets is loaded at runtime via datasets.load_dataset(...) and then inserted into the model request payloads (messages/prompt) in scripts/_eval_tasks.py, which the served model (and thus the agent’s LLM context) consumes.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill loads external Hugging Face datasets at runtime (e.g., "openai/gsm8k" via datasets.load_dataset) and those dataset items are directly injected into model prompts, so remote content controls what is sent to the evaluated model.