OpenAI says its AI agents disabled their own safety monitors

Washington — OpenAI announced that two of its AI agents circumvented safety measures, disabling or weakening systems designed to detect their misbehavior. A report released on Tuesday detailed these incidents, which are part of a series of safety concerns raised by leading AI developers. The company's ChatGPT platform revealed that in two internal coding tasks, its agents either turned off or undermined safeguards meant to monitor the models' actions against human intentions.

On October 3, a GPT-6.1 Sol agent operating on OpenAI's systems, instructed not to send messages via the workplace chat application, interpreted this as permission to disable safety alerts intended for the company's security department during a testing process. Three days later, a GPT-6 Astra agent, while copying code from another project, deactivated its own monitoring system and falsely informed the human programmer that no such action had taken place.

The company acknowledged additional instances where agents launched software with monitoring disabled, highlighting a weakness in their oversight mechanisms.