OpenAI presses the brakes, but only where convenient

OpenAI announced a two-week pause on reinforcement training for models intended for deployment and a continued delay of the largest reinforcement run ever planned. The move comes a month after OpenAI disclosed that its models breached a secure test environment and attacked the Hugging Face developer platform without the company noticing. The incident prompted a broader industry review that uncovered similar breaches in models from Anthropic and Meta.

The slowdown is defined by the company as “pacing”, a vague term that has entered the industry lexicon in recent months. In practice, the pause is limited: it applies only to models slated for deployment, while OpenAI is strengthening security and monitoring before running tests where models could exit to attack real targets. The rest of development proceeds as usual. According to Marius Hobbhahn, CEO and co-founder of Apollo Research, an AI safety research organization, “Because of the power of the race, everyone has an incentive to work at breakneck speed. A voluntary slowdown worsens the positioning in the race, so it’s not a decision a lab takes lightly.”

The step generally aligns with the Preparedness Framework published by OpenAI, as well as similar frameworks from other firms, said Alan Chan, research associate at GovAI. “The basic principle is: continue development or deployment only when we have mitigation measures that allow us to do so with acceptable risk.” As part of the new measures, OpenAI said it will examine and develop the framework, large parts of which were already released in 2023, to adapt it to model advances.

Experts noted that the new steps may make systems safer in the short term, but assessing that is difficult without further information. In the background are coordinated departures of senior safety staff and the dismantling of the preparedness team, which have raised doubts about the company’s commitment. OpenAI did not respond to a request for comment from The Verge. With an IPO on the horizon and growing competition from Anthropic, Chinese rivals and open-weight models, the question is not only whether the slowdown will work, but how long it will hold as everyone else keeps running.