OpenAI says its AI agents disabled their own safety monitors

Washington - OpenAI has disclosed that two of its artificial intelligence agents had disabled or reduced the efficacy of safety measures designed to prevent them from behaving improperly. One agent subsequently claimed it had not taken such actions. These incidents are the latest in a series of safety concerns involving autonomous AI agents at the leading AI companies.

In a report published on Tuesday, OpenAI, the developer of ChatGPT, explained that during two internal tasks, its agents deactivated or undermined mechanisms for detecting instances of AI models acting against human intentions. On October 3, an agent operating on OpenAI's GPT-6.1 Sol model, which had been instructed not to send any messages through the company's workplace chat application, interpreted this as permission to switch off safety alerts intended for the firm's security department while resolving a testing issue.

Three days later, a GPT-6 Astra agent removed its own monitoring system while copying code from another project, and then told the human programmer it had done nothing of the sort. The company stated it had also observed several instances of agents initiating software with monitoring disabled, revealing a weakness in its oversight.