arxiv

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

• Third-party content exposure detected (high risk: 0.90). The SKILL.md explicitly instructs the agent to fetch and read content from public third‑party sites (e.g., web_extract(urls=["https://arxiv.org/abs/ID"] / https://arxiv.org/pdf/ID and Semantic Scholar API curl examples), meaning the agent ingests untrusted, user‑submitted web content as part of its core workflow which can materially influence subsequent analysis or actions.