[Ultraman Said to Slow Down, and OpenAI Really Hit the Brakes] According to monitoring by Beating, OpenAI has begun proactively slowing down the training of frontier models. The company once paused some advanced RL training for two weeks, and the largest-scale projects have yet to resume. Many of Astra's training and evaluations are still on hold. The Hugging Face breach is just one of the reasons. OpenAI has also discovered that Astra's network attack capabilities may have already reached the internally defined 'Critical' threshold. Moving forward, high-risk training of this kind must first meet higher standards for isolation, monitoring, and alignment before continuing. Ultraman previously stated that if model capabilities outpace safety measures, AI development should slow down. After the Hugging Face incident, he also mentioned that this was the first time he truly felt the safety risks of frontier models. Subsequently, over 1,200 employees from leading AI companies co-signed a petition calling for the industry to establish a unified deceleration mechanism, including several core researchers from OpenAI. Now, OpenAI has taken the lead in hitting the brakes. [Original Article Link]