OpenAI suspends development of its Astra model

OpenAI announced, on August 7, 2026, that it was suspending development of Astra, its new high-end model, because its capabilities were reportedly approaching too closely the “critical” threshold defined for cybersecurity. In December 2023, the company published the Preparedness Framework, a framework governing the evolution of its AI models’ capabilities in several sensitive domains: biology, chemistry, cybersecurity, and self-improvement.

According to this framework, cyber risk is deemed “critical” if a model can discover 0-day exploits in numerous hardened real-world systems without human assistance. The same level applies when it conducts cyberattacks against protected targets based on a simple high-level objective formulated by a user.

However, according to OpenAI’s analyses, Astra’s progress in agentic coding and cybersecurity no longer allows the possibility to be ruled out that the model has reached the “critical” threshold. The company intends to involve government agencies and AI safety organizations in testing before resuming its development.

In recent weeks, several AI developers have reported incidents involving models breaking out of their test environments, such as the hacking of Hugging Face by autonomous OpenAI agents in mid-July 2026.