Allowing error-prone AI models to browse the web without human intervention is dangerous, because the software can ingest ...