OpenAI pauses Astra after it hits a 'Critical' cyber threshold
The first time a lab's own safety gate has delayed a frontier model: OpenAI's next flagship has no ship date. Plus Kimi K3 lands in GitHub Copilot.
Copy markdown
Critical, for the first time
In internal tests Astra could autonomously find and exploit well-defended, real-world systems. OpenAI says it "cannot rule out Critical capability level" — the top tier of its Preparedness Framework — making Astra the first model to trip that cyber bar.
What OpenAI did about it
OpenAI paused Astra work that doesn't meet stricter new guardrails, tightened its security controls, and pulled in government agencies and outside safety groups to stress-test the model before any release.
What it means for your roadmap
Astra was shaping up as OpenAI's next flagship and now has no ship date, so don't plan launches around it. Expect more frontier models to arrive behind cyber-capability reviews from here.
Elsewhere: Kimi K3 lands in Copilot
Moonshot's Kimi K3 is now a selectable model in GitHub Copilot, billed usage-based at provider list pricing — a one-click way to run its 1M-token, open-weights reasoning alongside Opus and GPT in your editor.