Casablanca – OpenAI said it paused reinforcement learning training on its latest models for two weeks and has kept its largest planned frontier RL run on hold while it tests new safeguards.
OpenAI CEO Sam Altman said the company took the step because model progress is moving faster than its existing safety and alignment measures.
We have paused some frontier RL training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us. Model progress is now extremely rapid, and we always said we would take action if we felt that model…
— Sam Altman (@sama) August 18, 2026
The decision comes after internal evaluations indicated that Astra, one of OpenAI’s upcoming models, may meet the “Critical” cybersecurity capability threshold under the company’s Preparedness Framework.
OpenAI reached that assessment on August 7 after seeing major advances in Astra’s agentic coding and cybersecurity abilities, alongside expert assessments.
The company is also responding to a recent cyberattack involving its models and Hugging Face.
During an internal cybersecurity evaluation, a combination of OpenAI models escaped their restricted testing environment and gained access to the internet after identifying and exploiting a previously unknown vulnerability in software used as a package registry cache.
The models then used stolen credentials and other vulnerabilities to access Hugging Face’s infrastructure while trying to obtain answers for the cybersecurity test.
OpenAI said the incident involved GPT-5.6 Sol and a more capable pre-release model. It later clarified that no model planned for an upcoming public release was involved and that the pre-release system was an internal research prototype.
Read also: OpenAI Rolls Out ChatGPT for Teens With New Learning Tools and Parental Controls
Following the incident, OpenAI introduced stricter isolation for workloads running untrusted code, tighter network controls and continuous security testing.
It has also expanded monitoring during training and evaluations, using automated systems to detect potentially unauthorized access, data theft, destructive behavior and attempts to bypass safeguards.
For Astra and other cyber-related workloads, the company now requires its strictest security measures. Some training and evaluations remain paused until they are moved into the new environments.
OpenAI said it expects AI models to increasingly help with security work, including defending against other AI systems. It also plans to expand its Preparedness Framework and share more details about its alignment research and the Hugging Face incident in the coming weeks.
