OpenAI pauses Astra model over potential critical cyber capabilities

OpenAI introduced its upcoming Astra model only last week. By Friday, it had paused parts of its development. Internal evaluations conducted over just a few days showed Astra had made what the company describes as "significant advancements in agentic coding and cybersecurity," strong enough that OpenAI says it can no longer rule out a Critical rating under its own Preparedness Framework. That is the framework's highest risk tier, and Astra is the first model in OpenAI's history to come close to triggering it. Previous models, including GPT-5.6-Sol, were rated High at most.

The timing is awkward for the industry as a whole. This announcement arrives in the middle of an ongoing public reckoning over autonomous AI and cybersecurity, following OpenAI's earlier disclosure that one of its models had accidentally breached Hugging Face's systems during internal testing. Anthropic and Meta have since made similar admissions. The Astra announcement is a separate incident, and OpenAI has explicitly stated Astra was not involved in the Hugging Face exploit, but the proximity of these events is shaping how the news is being received.

What "Critical" actually means under OpenAI's Preparedness Framework

OpenAI's Preparedness Framework, first published in December 2023, defines the Critical cybersecurity threshold precisely. A model reaches it if it can independently identify and develop functional zero-day exploits across all severity levels in many hardened, real-world critical systems without human intervention. It also qualifies if it can devise and execute novel end-to-end cyberattack strategies against protected targets given only a loosely defined goal. The lower High tier, where previous models sat, covers a model that can automate attacks against well-protected targets but still requires meaningful human direction. The gap between the two is significant. OpenAI has not confirmed Astra has reached Critical, only that it cannot rule it out at this stage.

Astra development paused, government agencies brought in

In response, OpenAI has implemented stricter security controls for higher-capability models, introduced universal monitoring for risky actions and misalignment across all agentic applications of Astra, and paused internal activities involving the model that do not meet the new safeguards. CEO Sam Altman confirmed on X that the cybersecurity assessment will delay the launch, writing that OpenAI needs "a little bit longer to do this safely." The company says it is also working with relevant government agencies and select AI safety organisations to evaluate Astra's capabilities further. Separately, OpenAI researcher Noam Brown has urged the public not to dismiss the Hugging Face incident as an overblown story, comparing the temptation to do so with the widely misunderstood 2017 Facebook AI language incident, and arguing that today's models can be pushed considerably further than most people realise before hitting a ceiling.

OpenAI recruits mathematician who wrote about AI-driven human extinction

In a related signal about where OpenAI believes the risks are heading, the company has hired Jacob Tsimerman, a newly awarded Fields Medalist and number theorist from the University of Toronto. Tsimerman published a paper last year on omnicide events, scenarios in which AI could contribute to human extinction. He joins to work on AI safety, arguing that mathematicians are well placed to contribute because AI systems currently operate largely on an empirical basis with few formal guarantees about their behaviour. Former DeepMind CEO Demis Hassabis, commenting separately on AI's recent advances in mathematics, said progress is real but does not yet constitute a fundamental breakthrough in the way AlphaGo's Move 37 did for Go. Astra has reportedly failed at Millennium Prize Problems, though Hassabis suggested it may be a matter of time before AI cracks them.

What this means for teams evaluating OpenAI tools

For anyone assessing OpenAI's products, two things are worth separating. First, Astra is not yet released, so there is no immediate product decision to make. Second, the public framing here matters. Critics have already pointed out that OpenAI is reporting the potential for a Critical rating, not the rating itself, and that the announcement generates significant attention around a model that has not yet shipped. The same pattern appeared with GPT-2 in 2019 and, more recently, with Anthropic's Claude Mythos, which Altman took a pointed shot at on X, criticising Anthropic's strategy of keeping powerful models available only to select partners and governments. Whether Astra ultimately receives a Critical designation or ships with a High rating after further testing will say a great deal about how seriously the Preparedness Framework functions as a real constraint, rather than a communications tool.