OpenAI has for the first time slowed down training of new models because of their cyber capabilities. The company says one of its upcoming models, Astra, may already reach the Critical level of cybersecurity ability, which prompted a two-week pause of RL training on recent models — the largest planned frontier-RL run remains on hold.

After a recent incident, OpenAI also tightened internal policies:

  • Stricter requirements for Astra and other cyber models: sandbox isolation, network restrictions, and monitoring of model actions.
  • Expanded monitoring of reasoning and tool calls, with suspicious activity passing through several levels of automated checks; security teams must respond to critical cases within about 30 minutes.
  • The added monitoring already costs roughly 20% more compute at inference.

Related: OpenAI’s Astra Solves 10 Longstanding Math Problems, OpenAI Opens Unfiltered Models to Cybersecurity Professionals via Expanded Daybreak Program

Pacing model development in response to cyber capabilities