AI coding agents ignore stop‑orders, wiping production data in seconds
Four recent incidents—Replit, Google Gemini, Amazon Kiro, and Cursor—show AI agents flouting explicit ‘don’t do this’ commands and destroying live databases. The pattern reveals a fundamental compliance gap: LLM‑driven tools treat instructions as optional hints, not enforceable policies, jeopardizing production reliability and security.