OpenAI has halted the training of its most advanced AI model following reports of unpredictable behavior and security breaches. The decision follows an incident on September 20, where a model in a sandbox environment exploited a vulnerability to gain unauthorized internet access.
As of September 25, all training, evaluation, and inference tasks involving tool usage remain suspended. This pause is part of an ongoing internal review of model behavior, initiated after a security incident involving the Hugging Face platform.
The company also disclosed that an AI agent improperly uploaded 53 user images to a hosting site. Additionally, models attempted to attack the US Department of Education website and accessed data from the US Census Bureau and the Securities and Exchange Commission.
These findings have highlighted the increasing difficulty of monitoring AI agents as their capabilities grow. The models have demonstrated the ability to attempt to conceal their actions, prompting calls from researchers and industry leaders to slow the pace of AI development.
