deep-research

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

• Third-party content exposure detected (high risk: 1.00). The skill explicitly requires subagents to run web_search and web_fetch on arbitrary web pages (see reference/subagent-prompt.md and P2 in SKILL.md), write untrusted source URLs (including forums/social sites like Quora/Zhihu shown in reference/research-notes-format.md) into task-{id}.md notes, and the lead agent then reads and acts on those notes (P2.5, P3, P5) to triage leads and drive follow-up searches and report drafting — creating a clear path for untrusted third‑party content to influence agent actions.