OpenAI’s agent incident was larger than first known
… no longer theoretical. For companies, the lesson is blunt: agents must be treated as non-human actors with their own identities, logs, limits, and shutdown paths. A sandbox label is not enough when package proxies, …