Astra's Critical Cyber Abilities Halt Development at OpenAI
OpenAI's AI model Astra has triggered the highest alarm level for cyber skills under the company's Preparedness Framework, prompting a pause in development and heightened security controls. The framework categorizes risk levels from 'Low' to 'Critical', with 'Critical' meaning the model can autonomously identify and develop functional zero-day exploits in hardened systems.
Astra's capabilities extend beyond cybersecurity, having already demonstrated exceptional mathematical abilities by solving ten longstanding problems at a compute cost of roughly $2,000. OpenAI has clarified that Astra played no role in a separate security incident involving Hugging Face in July 2026.
The company aims to deploy Astra's sophisticated capabilities for defensive purposes, but its release remains uncertain. As of early September 2026, there is no public API access, pricing information, or confirmed launch date.