tao-finetune-cosmos-reason

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

• Third-party content exposure detected (high risk: 0.75). At runtime, the evaluate action ingests outsider-authored free text from the user-provided evaluation dataset annotation JSON (dataset.annotation_path → cosmos-rl-evaluate --config {config_path}), where fields like question/gt are read and used as prompt/context for the model.

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

• Potentially malicious external URL detected (high risk: 0.90). The packaged conversion helper (scripts/prepare_cosmos3_vlm_checkpoint.py) may git-clone and then execute code from the remote repository https://github.com/NVIDIA/cosmos-framework.git (and pulls/executes the conversion container image nvcr.io/nvidia/pytorch:25.09-py3) at runtime, which fetches and runs external code; therefore this is a runtime external dependency that can execute remote code.