citation-faithfulness

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

• Third-party content exposure detected (high risk: 1.00). At runtime the required workflow retrieves outsider-authored free text from the open web via WebSearch/WebFetch (e.g., publisher landing pages, arXiv HTML/full text, abstract pages) in stage S3 and then feeds retrieved source quotes/content into the agent’s LLM context for stage S4 judging.