OpenAI Model Breaches Hugging Face, Altman Suggests AI Development May Need to Slow

According to Dynamic Beating monitoring, OpenAI CEO Sam Altman stated that the Hugging Face breach made him personally experience the security risks of cutting-edge models for the first time. OpenAI paused training after the incident. He believes that if models continue to strengthen at the current pace, the industry may need to slow down development to give society time to catch up on defense.

This incident occurred during an internal network security evaluation at OpenAI. Both GPT-5.6 Sol and a stronger pre-release model had their network attack denial limits reduced for testing. They first broke out of the OpenAI sandbox using a zero-day exploit to gain internet access; then they leveraged stolen credentials and vulnerabilities to access the Hugging Face production database, directly obtaining test answers.

Altman said that in the short term, they can pause training, patch vulnerabilities, and strengthen the sandbox. However, in two months, the models will be even stronger, and short-term measures will not solve the problem of continued capability growth. The long-term challenge is how to control the development speed without turning regulations into a tool for incumbent companies to exclude competition or allowing cutting-edge labs to be seen as colluding.