Washington: OpenAI has temporarily paused some frontier reinforcement learning (RL) training as rapid advances in AI capabilities raise concerns that its safety, alignment, security and monitoring measures may not be keeping pace.
OpenAI CEO Sam Altman said the company had taken the step because model development had accelerated significantly.
"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," Altman said in a social media post.
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…
"Model progress is now extremely rapid, and we always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment," he added.
OpenAI said two recent developments had highlighted the growing risks associated with increasingly capable AI systems: an incident involving OpenAI and Hugging Face, and preliminary evidence that one of its upcoming models, Astra, may meet the "Critical cybersecurity capability" threshold under its Preparedness Framework.
The company said it had therefore temporarily slowed the pace of scaling, including a two-week pause in RL training on its latest models intended for deployment.
During the pause, OpenAI said it had further hardened and red-teamed its research environments and expanded monitoring coverage. Its largest planned frontier RL run remains on hold while smaller-scale training and evaluations continue.
The company said the additional work is intended to assess model behaviour, validate safeguards and gather stronger evidence of alignment before the larger training run resumes.
Stronger security measures
OpenAI said increasingly capable frontier models also require stronger security measures in the environments where they are trained and evaluated.
The company has introduced tighter workload and network isolation, stronger security testing and improved monitoring of research environments. It said some workloads involving Astra or cyber-related models remain paused until they meet the new security requirements.
OpenAI's Preparedness Framework treats certain critical cybersecurity capabilities as requiring safeguards during development, reflecting the potential for highly capable models to identify and execute sophisticated cyberattacks.
The company has also expanded monitoring of frontier RL training and evaluations involving tools. OpenAI said its new monitoring system can examine model activity and tool use for potential unauthorised access, data theft, destructive behaviour and attempts to circumvent safeguards.
Alignment becomes a bigger focus
OpenAI said it is applying core alignment techniques across more stages of training for its most capable models.
These measures include improving reward models to identify unsafe behaviour, training models to be more transparent about their actions and limitations, and reducing behaviours that exploit weaknesses in rewards, graders, tools or human oversight.
Altman said AI safety would play an increasingly important role in determining the pace of future model development.
He also called for greater coordination across the AI industry around common safety standards.
"We believe the entire field will have to coordinate on shared safety standards, but will act unilaterally in the meantime," Altman said.
Despite the temporary pause, OpenAI said it remained committed to developing and making advanced AI capabilities widely available.
The company said the latest measures are intended to ensure that its ability to understand, align and secure increasingly capable models keeps pace with their development.