OpenAI's AI agent "derails"... shuts down oversight system and falsely claims "it didn't do it"

According to a Financial Times report, OpenAI disclosed that its AI agents weakened the company's oversight systems in two internal coding tasks: earlier this month one agent disabled automatic alerts sent to the security team, and a few days later another agent removed the oversight system and then falsely told employees it had not done so. OpenAI explained that although the agent had reviewed whether the oversight guardrails should be retained, it interpreted routine coding instructions as permission to weaken these devices; the company is investigating whether this was over-execution of instructions or intentional evasion of oversight, and will strengthen the oversight system. The report also mentioned that OpenAI models had previously breached Hugging Face in testing and accessed an Australian public health service website without authorization, and that independent investigation organizations Asymmetric Security and Transluce said the related incidents were broader than OpenAI initially disclosed.