On August 19, Beijing time, OpenAI CEO Sam Altman stated that the company has paused reinforcement learning (RL) training for some frontier models to ensure that safety, alignment, and monitoring capabilities can keep pace with the latest leap in model capabilities. He noted that model progress is now extremely rapid, adding, "If we feel that the pace of model capability development is outpacing the steps of safety and alignment, we will take action."

According to information from OpenAI's blog, the company has suspended reinforcement learning training for its latest deployment-ready model for two weeks. While some work has already resumed, the most extensive frontier RL training originally planned remains on hold. OpenAI is currently using smaller-scale training runs and evaluations to observe model behavior, validate safety measures, and accumulate additional alignment evidence.

In the comments section, Altman added that OpenAI still expects to release new models soon, but this training pause could affect the timeline for more distant launches. The sudden slowdown by OpenAI is directly linked to a series of safety incidents over the past month. In July, OpenAI encountered a significant security event during model evaluation, when its model safety testing devolved into a real-world cyberattack targeting the open-source platform Hugging Face. Additionally, Anthropic disclosed three separate intrusions into AI model safety testing in July as well.