OpenAI Pauses Astra After It Couldn't Rule Out 'Critical' Cyber Capabilities

For the first time, OpenAI publicly slowed a model's development because it could not exclude the possibility that the upcoming Astra model has Critical-level offensive cybersecurity abilities, including autonomous zero-day potential.

OpenAI acknowledged a "Critical" cyber threat classification on its upcoming Astra model, saying it "could not be ruled out as having Critical-level cybersecurity capabilities," according to a report surfaced by @gh523516. The company responded by adding isolated testing procedures and slowing the model's rollout — the first time OpenAI has publicly braked development over loss-of-control concerns rather than shipping and patching.

Unlock the full briefing

Get every story in today's briefing, the full archive, and the daily AI intelligence brief.

All stories today

Full archive

Daily brief

Cancel anytime. Payments powered by Stripe.