tao-train-mask2former

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

• Third-party content exposure detected (medium risk: 0.30). SKILL.md specifies that the runtime train action requires user-provided dataset spec values (e.g., label maps and COCO panoptic/instance JSON files under dataset.*_json) which are LLM-ingested as part of the structured config and then read/parsed by the underlying job (outsider-controlled file contents can therefore include free text).

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

• Potentially malicious external URL detected (high risk: 0.90). The deploy Quick Start includes docker run commands that pull and run the external container image nvcr.io/nvidia/tao/tao-toolkit:6.26.3-deploy (e.g., "docker run ... nvcr.io/nvidia/tao/tao-toolkit:6.26.3-deploy"), which will fetch and execute remote code at runtime.