OpenAI's Astra is its first 'Critical'-risk AI model

OpenAI built a new model called Astra. It isn't out yet. On September 1, OpenAI said Astra is the first model it plans to release. It crosses OpenAI's own 'Critical' bar for cybersecurity danger. That means it's good enough, in testing, to pull off serious, novel cyberattacks. Nothing OpenAI has built before hit that level. The company is now restricting Astra's release and limiting who gets its cyber abilities.

Astra wasn't involved in the cyberattack that hit Hugging Face, the AI code-sharing site, this past summer. But that breach still changed OpenAI's plans. The company paused parts of Astra's development afterward, to add stronger protections against exactly this kind of misuse. There's an important caveat here too. 'Critical' isn't a label some outside safety regulator applied. It's OpenAI grading its own model, using its own internal rulebook. No independent body has confirmed the rating. Critics of AI safety self-policing will likely point to that.

Keep reading — what to watch

Will OpenAI actually release Astra's full cyber capabilities to anyone outside a small, vetted group? The company hasn't said exactly who gets access, or when broader access might follow. That answer will say a lot about how seriously AI labs take their own warnings. It's also worth watching whether other AI labs — Anthropic, Google DeepMind — start reporting models crossing this same bar. If Astra isn't the last model to cross it, that's a new kind of threat. Cybersecurity teams everywhere will need to adjust.