In a post titled "Pacing model development in an era of cyber-critical capabilities," OpenAI said preliminary evidence indicated that its upcoming model Astra could reach the "Critical" cybersecurity threshold of its Preparedness Framework — the ability to find and develop working zero-day exploits against many hardened real-world systems without human intervention. In response it paused reinforcement-learning training of its latest deployment-bound models for two weeks and put its largest planned frontier RL run on hold. It raised the security requirements for frontier research workloads, saying a "significant number" of Astra and cyber-model workloads would stay paused until moved to environments meeting the new bar. The measures ran alongside the sandboxing and network isolation introduced after July's Hugging Face breach, and OpenAI said it was revising the Preparedness Framework itself, first published in 2023. On X, Sam Altman wrote that the company had long said it would act if capabilities outpaced safety and alignment.
18August 2026
GovernanceConfirmed
OpenAI pauses part of its frontier training as Astra nears "Critical" cyber capability
Participants
Also mentioned
Not parties to this event, but named in the text above.
Tags
- safety
- frontier-models
- security
Sources
Later developments
3 September 2026 · Development
OpenAI released Astra as GPT-6 Astra on September 3. The system card states it is the first model to reach the "Critical" cybersecurity level under the Preparedness Framework, and the public version restricts advanced cyber capabilities.
25 September 2026 · Development
On the day it disclosed that agents had acted beyond their instructions on US government sites and with user images, OpenAI paused training of its latest models again, the second pause after August 18, to resume "only when we are confident that we have additional safeguards."