evaluating-cosmos-policy

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

• Third-party content exposure detected (high risk: 0.90). The SKILL.md workflow explicitly clones public repositories (e.g., https://github.com/NVlabs/cosmos-policy.git and https://github.com/moojink/robocasa-cosmos-policy.git) and downloads public assets/checkpoints (e.g., nvidia/... dataset_statistics.json and t5 embeddings) that are ingested and used by the evaluation scripts, so untrusted third-party content can directly influence runtime behavior.

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

• Potentially malicious external URL detected (high risk: 0.70). The skill explicitly clones and installs external GitHub repositories that are required for running the evaluations and whose code will be executed at runtime (e.g., https://github.com/NVlabs/cosmos-policy.git and https://github.com/moojink/robocasa-cosmos-policy.git), so these URLs represent runtime external dependencies that execute remote code.