mantis-architecture

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

• Third-party content exposure detected (high risk: 0.70). SKILL.md:23-28 requires parsing learnings.jsonl / historical_learnings.jsonl and extracting “trajectory insights” and “discovered vulnerabilities,” which are typically outsider-authored free-form text (e.g., prior agent/user messages) that then gets synthesized into Markdown and fed back into the agent’s LLM context during KB construction.