deep-research

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

• Third-party content exposure detected (high risk: 0.90). SKILL.md Step 2–3 requires running runtime web search / fetching (e.g., WebSearch, WebFetch, agent-browser, and /last30days which includes Reddit/X/YouTube/HN text), whose retrieved free-form page/message content is then summarized and used in Step 4 synthesis, meaning outsider-authored prose is ingested into the agent’s LLM context.