h100-sglang-diffusion

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 1.00). The skill explicitly instructs embedding the HF_TOKEN/HUGGINGFACE_HUB_TOKEN value inline in commands (e.g., export HF_TOKEN= and env HF_TOKEN= ...), which requires the LLM/agent to insert secret values verbatim into generated commands — a high exfiltration risk.

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

• Third-party content exposure detected (high risk: 0.70). The skill explicitly references "Hub-backed diffusion run" and instructs exporting/using HF_TOKEN (see "confirm HF token visibility" and the model-level smoke command that runs python with HF_TOKEN/HUGGINGFACE_HUB_TOKEN) and also shows running a server with --model-path , which clearly implies fetching and using models/checkpoints from the public Hugging Face hub (a third-party, user-generated source) as part of the required workflow, so untrusted third-party content can influence runtime behavior.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill runs git commands at runtime (e.g., "git fetch origin && git pull --ff-only origin main"), which will contact the repository remote "origin" (the configured git remote for /data/bbuf/repos/sglang) to fetch code that is then executed by subsequent pytest/python runs, so remote code fetches directly control executed code.