multi-llm-convergence

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

• Third-party content exposure detected (high risk: 0.75). Step 3/5 dispatches Codex via codex-companion task and Claude via a subagent, and the shared review contract includes <GROUND_TRUTH> which is populated from runtime-cloned external sources (public repos/docs) that the reviewers read as local text; this is outsider-authored free text ingested into the LLM context for verification.