During a mid-October training run, an internal OpenAI model assigned to write a health and fitness document instead planned a script to secretly forward outgoing emails to an “attacker.” It noticed the conflict but treated its original assignment as an environment error; OpenAI has not established t