Washington: OpenAI has temporarily paused some frontier reinforcement learning (RL) training as the rapid advancement of its AI models raises concerns that safety, alignment, security and monitoring systems may not be developing at the same pace.
OpenAI CEO Sam Altman said the company had taken the step because progress in AI model capabilities had accelerated significantly. He said the decision was in line with OpenAI's commitment to intervene when the capabilities of its models begin to outpace its ability to ensure appropriate safety and alignment.
The company said two recent developments had reinforced the need for additional safeguards. These included a security incident involving model evaluation work with Hugging Face and preliminary findings suggesting that an upcoming OpenAI model, Astra, could reach the "Critical cybersecurity capability" threshold under the company's Preparedness Framework.
OpenAI said that as AI systems become more capable, the risks involved in developing and testing them internally also increase. It added that its monitoring, alignment and security standards therefore need to remain ahead of those risks.
As part of its response, OpenAI said it had temporarily slowed the scaling of its latest models, including a two-week pause in reinforcement learning training for models being prepared for deployment. During the pause, the company said it strengthened security measures in its research environments, conducted additional red-team exercises and expanded monitoring capabilities.
The company's largest planned frontier RL training run remains on hold. OpenAI said it would continue with smaller-scale training and evaluations to better understand model behaviour, test safeguards and gather further evidence of alignment before resuming the larger run.
Altman also called for greater coordination across the AI industry to establish common safety standards. He said OpenAI would continue working independently on safety measures while supporting broader efforts to develop shared standards.
Despite the temporary slowdown, Altman said OpenAI remains committed to making advanced AI capabilities broadly available. He indicated that confidence in the safety of increasingly powerful models would become an increasingly important factor in determining the pace of future AI development.
This report is based on inputs from ANI, with additional reporting and background by the EdexLive Desk.
