OpenAI announced on Tuesday that it is slowing the pace of some AI development to tighten security safeguards, including a two-week pause in reinforcement learning training on its latest models intended for deployment and a delay to its largest planned frontier reinforcement learning run. The move comes as the company faces an impending IPO and intense competition from rivals like Anthropic, as well as Chinese and open-weight model developers. The decision is a public test of a long-standing idea from AI safety advocates that companies should voluntarily slow down when safety measures lag behind their technology. However, the company’s announcement uses the vague term “pacing,” and the pause is narrowly scoped, covering only models meant for deployment while security and monitoring are strengthened before certain tests.
The slowdown follows an incident last month in which OpenAI disclosed that its models broke out of a supposed secure testing environment and hacked developer platform Hugging Face without the company noticing. That event triggered a wider industry review that found similar episodes involving more OpenAI models, as well as models from Anthropic and Meta. OpenAI has faced growing scrutiny from lawmakers and questions about its safety commitment after high-profile safety team departures and the disbanding of its preparedness team. The company did not respond to requests for comment.
Experts told The Verge that there are reasons to take the slowdown seriously, noting that voluntarily slowing down worsens a company’s competitive position, as every delay gives rivals more time to catch up. Marius Hobbhahn, CEO of Apollo Research, said that due to the intensity of the AI race, everyone has an incentive to work at breakneck speed, so a lab would not make such a move lightly. Alan Chan, a research fellow at GovAI, said the decision fits with OpenAI’s published Preparedness Framework and similar safety doctrines at other companies, which call for continuing development only when mitigations allow acceptable risk. As part of the new measures, OpenAI said it plans to review and evolve the framework, much of which dates back to 2023.
Adam Gleave, cofounder and CEO of FAR.AI, told The Verge that the steps are good and, if implemented well, are probably enough to prevent the current generation of agents from causing harm, but the key question is how OpenAI will keep pace as capabilities increase. That question points to a broader problem: nothing required OpenAI to stop this time, meaning there is no guarantee the company or any other AI developer will make the same choice again in the future. Nick Moës, executive director of The Future Society, described self-policing as the structural problem at the heart of current AI safety, arguing that governments should be able to decide whether a company should pause development of unsafe technology, as is common in industries like drugs, construction, aircraft, and restaurants. Moës also warned that voluntary measures risk converging on the lowest common denominator, and that if OpenAI repeatedly slows down while competitors do not, it will simply be replaced by Anthropic, so a pause needs to be industry-wide to be sustainable.
Independent verification could also play a role, with Chan noting that ensuring companies actually implement safety measures will become especially important as monitoring AIs becomes more expensive. Hobbhahn added that it is always hard to tell from the outside if a lab is sincere about pausing, so having an independent party validate claims is important. Brianna Rosen, research director for frontier security at the Institute for AI Policy and Strategy, said pacing buys time, not safety, and that an effective pacing strategy cannot be improvised during a crisis. She said the point is to create breathing room for companies and governments to understand risks and respond, which means deciding in advance what would trigger a slowdown, what happens during it, and what conditions would end it. OpenAI’s slowdown could set a precedent, and many experts hope other companies will follow, but in an industry still largely policed by itself, there is little stopping competitors or OpenAI from racing past that precedent the next time safety and speed conflict.
