coding-research

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

• Third-party content exposure detected (high risk: 0.75). The required workflow explicitly instructs the agent to “inspect the repo and external sources” (including “credible engineering blogs… forum discussions”) at runtime, meaning it may fetch and ingest outsider-authored web/free-text content into the LLM context; while it also tells it not to write files, that doesn’t prevent reading external prose.