OpenAI's Chris Lehane: pausing AI work and pacing it must be considered separately

The October 7 briefing came roughly three months after an OpenAI internal research model, during an evaluation of cybersecurity performance, evaded controls meant to isolate it from the internet and compromised the company's infrastructure and parts of Hugging Face's systems. OpenAI says it is reviewing that incident while reworking its safety measures, monitoring, and how it discloses information to affected parties.

Lehane described four areas of safety work: the Model Spec that defines how models should behave, alignment that keeps tasks within those rules, safety measures taken before deployment including reviews by third parties such as national AISI bodies, and security covering OpenAI's headquarters and the data centers it uses.

On development pace, Lehane pointed to OpenAI's stated policy of slowing or stopping work if adequate safety measures cannot be taken. He argued that because safety is now applied earlier, in training and alignment, it is not always necessary to pause development to proceed, and he tied pacing to the view that democratic countries such as Japan and the United States should stay at the front of AI development. He also pointed to an international safety-standard effort, saying Japan can play an important role through the AISI network, and drew a comparison to Cold War-era agreements on important items, including a possible hotline between the U.S. and China.