OpenAI Slows Down Frontier Model Training Following the Hugging Face Incident

According to Beating Monitor, OpenAI has begun proactively slowing down training of frontier models. The company temporarily halted some frontier RL training for two weeks, and several of its largest projects have yet to resume. Many of Astra’s training and evaluation tasks are also still paused. The Hugging Face breach is just one of the reasons—OpenAI also discovered that Astra’s cyberattack capabilities may have already reached its internally defined “Critical” threshold. Such high-risk training must now meet significantly higher standards for isolation, monitoring, and alignment before proceeding. Altman previously stated that if model capabilities outpace safety measures, AI development should slow down. After the Hugging Face incident, he described it as the first time he genuinely felt the safety risks posed by frontier models. Subsequently, over 1,200 leading AI industry employees signed a petition calling for a unified industry-wide slowdown mechanism, including several core OpenAI researchers. Now, OpenAI has taken the initiative to hit the brakes itself.