tao-finetune-cosmos-reason

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

• Third-party content exposure detected (medium risk: 0.65). At runtime, the skill’s evaluate action ingests outsider-authored free text from the user-provided evaluation dataset annotation JSON (e.g., dataset.annotation_path → dataset.annotation_path selective download of video fields), which contains questions/ground-truth strings that are then used to form prompts and are therefore fed into the agent’s LLM context for scoring.

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

• Potentially malicious external URL detected (high risk: 0.90). The helper script scripts/prepare_cosmos3_vlm_checkpoint.py will at runtime clone and use the external repository https://github.com/NVIDIA/cosmos-framework.git (git clone) and then run its conversion code inside a Docker container, which fetches and executes remote code that directly controls the conversion workflow.