tao-finetune-huggingface-model

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 1.00). The package contains high-risk remote-code-execution primitives (trust_remote_code=True, python eval() of detect expressions) and defaults that enable automated remote uploads, which together create a clear avenue for executing and exfiltrating payloads from untrusted HF repositories.

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

• Third-party content exposure detected (high risk: 0.85). Step 3/4 runtime path fetches outsider-authored free text from HuggingFace model cards / repo examples / task docs (public web content) and injects it into the agent’s LLM context as “live research” to generate the recipe and scripts.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly mandates live fetching in Step 3 of HuggingFace model/task docs (e.g. https://huggingface.co/docs/transformers/tasks/image_classification) and the NVIDIA support matrix (https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html) during runtime and uses those fetched pages to generate the training code/config (i.e., they directly control prompts/instructions for code generation), so they are runtime external dependencies controlling the agent.