ml-research

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

• Third-party content exposure detected (high risk: 0.90). The SKILL.md mandates a Web mode that requires making multiple WebFetch calls to public third‑party sources (official docs, GitHub READMEs/examples, HuggingFace model cards, and even GitHub issues/discussions) and then reading and grounding responses on that fetched content, which exposes the agent to untrusted, user‑generated web content that can influence subsequent tool use and decisions.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly requires runtime WebFetch of external docs and GitHub content to ground and control agent instructions (e.g., https://github.com/arcee-ai/mergekit is listed in the Web mode URL registry), so fetched remote content directly influences prompts at runtime.