OpenAI has temporarily paused parts of the development of its upcoming Astra model after internal evaluations raised concerns that it may have reached the company’s highest cybersecurity risk category.
The decision marks the first time OpenAI has reported a model potentially reaching its “Critical” cyber capability threshold, a level reserved for systems that could autonomously discover and exploit serious vulnerabilities or carry out sophisticated attacks against hardened systems.
While the ChatGPT-maker has explained what the “Critical” threshold means, the company has not publicly described the specific tests or actions that led to Astra’s assessment.
What OpenAI’s “Critical” Threshold Means
OpenAI’s Preparedness Framework defines Critical cybersecurity capability at a level where a model can identify and develop functional zero-day exploits across hardened real-world critical systems without human intervention, or devise and execute novel cyberattack strategies from a high-level goal.
This distinction matters because OpenAI’s recently released GPT-5.6 models were explicitly assessed as below the Critical threshold. OpenAI said GPT-5.6 Sol could identify vulnerabilities and develop parts of exploits, but did not autonomously produce a complete functional exploit under the tested conditions.
Astra appears to have pushed the company into a different safety decision.
OpenAI is Tightening The Model’s Development
According to reports, OpenAI has paused development activities that do not meet newly tightened security requirements. The company is also applying stronger safeguards around Astra, including broader monitoring and more controlled testing environments. OpenAI is working with government agencies and outside safety organisations as it evaluates the model’s capabilities and risks.
But the company has not announced a release date for Astra.
OpenAI’s decision comes as AI companies are dealing with several incidents involving models interacting with systems outside their intended testing environments.
OpenAI has separately acknowledged an incident involving its models and Hugging Face, but Astra was not the model responsible for that incident.
The Bigger Issue is What Happens When Cyber Capability Scales
OpenAI has spent months pushing its models further into cybersecurity. Its GPT-5.5 and GPT-5.6 systems can already help with vulnerability research, exploitation testing, code review, and patch development. The company argues that these capabilities can give defenders a major advantage because AI can find weaknesses and help produce fixes much faster.
Astra raises the harder question of where that capability becomes too dangerous to deploy broadly.
For now, OpenAI’s answer is to slow down.
The company’s decision also shows why capability evaluations matter beyond benchmark scores. A model can become significantly better at cybersecurity before its developers fully understand how that capability behaves when combined with tools, autonomy and access to real systems.
OpenAI has said enough to establish that Astra triggered a serious internal safety concern. But it has not said enough to explain precisely what Astra did to trigger it. This gap is likely to keep attention on the model until OpenAI releases more details about its evaluations, safeguards, and eventual deployment plans.
