multi-node-slurm

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

• Third-party content exposure detected (high risk: 1.00). The skill's workflow (SKILL.md) explicitly downloads and uses remote model artifacts and custom modeling code from public hubs (e.g., via HF_TOKEN / trust_remote_code=True which triggers save_artifacts to download tokenizer/config/custom model code from Hugging Face and possibly GitHub), meaning untrusted third-party content is fetched and can execute or alter runtime behavior.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill instructs at runtime to load Hugging Face model artifacts (e.g., --hf_model_path / → https://huggingface.co//), and explicitly warns that with trust_remote_code=True rank 0 downloads tokenizer/config/custom modeling code during save_artifacts — i.e., fetching and executing remote model code at runtime which the job depends on.