mindspeed-mm-vlm

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

• Third-party content exposure detected (high risk: 0.90). The skill explicitly downloads and ingests public third-party content required for training—e.g., "huggingface-cli download" for HF weights in Step 1.1, downloading LLaVA-Instruct-150K and COCO in Step 2, and cloning external repos in Step P1—which the workflow processes as training data/weights and can materially change model behavior.

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

• Potentially malicious external URL detected (high risk: 1.00). The skill explicitly runs setup and runtime commands that clone and install remote code (e.g., git clone https://gitcode.com/Ascend/MindSpeed-MM.git, git clone https://github.com/NVIDIA/Megatron-LM.git, git clone https://gitcode.com/Ascend/MindSpeed.git, pip-install wheels from https://gitcode.com/Ascend/pytorch/releases and git clone https://github.com/inclusionAI/Ming.git), which are fetched during runtime and will install/execute remote code required for the skill.