add-benchmark

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

• Third-party content exposure detected (high risk: 0.80). The skill explicitly downloads and ingests external datasets via ng_prepare_data from the GitLab dataset registry and instructs wrapping/calling third‑party benchmark libraries (see "Step 2: Prepare data" and "External Benchmark Integration"), and those untrusted/user‑provided artifacts (JSONL/verifier_metadata and external library outputs) are parsed by verify()/agent run logic and directly influence test execution, rewards, and follow‑up actions.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill requires runtime dataset upload/download via the MLflow/GitLab tracking URI (e.g. https:///api/v4/projects/<PROJECT_ID>/ml/mlflow), which is used to fetch JSONL datasets that directly control prompts and is a required dependency for train/validation workflows.