deep-research

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

• Third-party content exposure detected (high risk: 0.75). The skill’s runtime workflow explicitly mandates that subagents write “full findings” including “verbatim source URLs” to durable files and then the orchestrator reads those files to synthesize; if any subagent’s research pulls public web content or other outsiders’ text, that outsider-authored free text is ingested into the LLM context during research and then persisted.