monkey-patch-kernels-to-transformers

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

• Third-party content exposure detected (high risk: 1.00). The skill's Research phase (references/auto-kernelize.md and references/kernel-integration.md) explicitly requires using WebSearch/WebFetch against public sources like Hugging Face model cards and GitHub and instructs the agent to read and act on that content to plan and execute patches, so untrusted third-party content can influence decisions and tool use.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly instructs using WebSearch/WebFetch against GitHub and the HuggingFace Hub (e.g., https://github.com and https://huggingface.co) during runtime to fetch model implementation files which may be loaded/executed (via trust_remote_code) and used to drive the agent's integration/patching logic, creating a runtime dependency that can execute remote code.