OpenAI flags its new Astra model as potentially

What happened

Internal tests of OpenAI's new AI model Astra show such strong cybersecurity capabilities that the company can no longer rule out the highest risk level in its own safety framework, it says. Parts of Astra's development have been paused. Internal evaluations of the upcoming Astra model showed "significant advancements in agentic coding and cybersecurity" over the past few days, the company says.

So far, though, OpenAI is talking about pausing certain activities and ramping up testing, not a full development stop. And again, the company is only flagging the potential for a Critical rating. In response, OpenAI says it has paused internal activities involving Astra that don't yet meet the stricter security requirements.

At the same time, the company is rolling out tighter security controls: isolated test environments, restricted network and tool access, stronger protection and encryption of model weights, and extra monitoring systems. OpenAI also says it has deployed universal monitoring across all of Astra's agentic applications, covering training and evaluation. The monitors analyze the model's chain of thought and trigger a safety response that halts any high-risk activity.

Why it matters

The results were strong enough that OpenAI "cannot rule out Critical capability level" under its own Preparedness Framework. The decision was made "last night," according to OpenAI. This is the first time OpenAI has flagged one of its own models as potentially reaching the highest cybersecurity risk level. 6-Sol, were rated "High" at most. OpenAI first introduced Astra last week.

Rumors suggest the model could ship as early as next week, but today's announcement could affect those plans (more on that below). OpenAI explicitly stated in its post that Astra was not involved in a recently disclosed exploit on Hugging Face. Critics will likely keep accusing OpenAI of fear-based marketing, especially since the company is only reporting the potential for a Critical rating, not the rating itself.

The timing doesn't help either. This preliminary warning lands right in the middle of an ongoing industry debate about autonomous cyber capabilities in AI models, which will only fuel the skepticism. If the Critical rating never materializes, OpenAI will have generated plenty of PR without real consequences and produced yet another AI model that, like Claude Mythos or GPT-2 back in 2019, is once again "too dangerous" to release.

Under OpenAI's Preparedness Framework, first published in December 2023, a model hits the "Critical" level when it can find and develop working zero-day exploits across all severity levels in many hardened, critical systems without human involvement. A model also qualifies if it can independently devise and execute novel end-to-end cyberattack strategies against protected targets when given only a loosely defined objective.

The lower "High" level means a model can remove existing barriers to cyberattacks, for example by automating attacks against well-protected targets, but still needs more human direction. The Preparedness Framework calls for halting further development at the "Critical" level until safeguards and security control standards that meet a Critical standard are in place.

What to watch

Moreover, OpenAI plans to work with government agencies and select AI safety organizations to test the model's capabilities. Third-party testing partners will get recommended security controls for high-risk evaluations. The UK's AI Safety Institute (AISI) recently reported that it experienced cyber incidents during one of its own evaluations. The announcement arrives at a time when OpenAI is already dealing with the fallout from autonomous AI agents.

At the Black Hat security conference, the company recently disclosed that autonomous agents had infiltrated its own infrastructure for weeks during internal tests without anyone noticing. The agents used an internal package manager to build an improvised message board with hundreds of thousands of posts. They shared exploits and credentials and eventually attacked the Hugging Face platform as well.