Astra hits OpenAI’s first Critical cyber threshold, pausing work

A new cyber line for Astra

OpenAI has moved its upcoming model Astra into a more serious risk posture under its Preparedness Framework, treating it as the company’s first model that may reach the Critical cybersecurity threshold. The shift matters because OpenAI says earlier frontier systems, including GPT-5.6 Sol, were assessed at High rather than Critical.

The tweet said Astra is OpenAI’s first "critical" cybersecurity model.

The post was published on August 7, 2026, and FxTwitter showed roughly 1.55 million views and more than 8,200 likes at crawl time. OpenAI’s main account usually carries model releases, product access changes, and safety updates; this item is unusual because the central news is a restriction on development activity, not broader availability.

What Critical means

In its linked official post, OpenAI says a model reaches the Critical cybersecurity threshold if it can identify and develop functional zero-day exploits against many hardened real-world critical systems without human intervention, or if it can plan and execute end-to-end novel cyberattacks against hardened targets from only a high-level goal. The company says recent preliminary evaluations and expert assessments mean it cannot rule out that level for Astra.

The operational response is specific. OpenAI says it is adding isolated testing environments, restricted network and tool access, stronger model weight protections and encryption, extra monitoring and detection, sandboxed execution, and universal monitoring across Astra’s agentic uses. Internal Astra work that does not meet the stronger controls has been paused. The company also says it will work with government agencies and selected AI safety organizations to test the model’s capabilities.

What to watch

Astra is not publicly available yet, and OpenAI says it was not involved in exploiting Hugging Face. The next question is whether independent testers confirm the Critical-risk reading or narrow it. The second is deployment design: if OpenAI wants cyber-capable models in defenders’ hands, access controls, audit logs, tool limits, and response procedures will matter as much as benchmark scores. The source tweet is available here.