OpenAI has built an AI model called Astra that can find zero-day vulnerabilities and chain them into working exploits without step-by-step human guidance. The model isn't public yet — access is limited to a small group of testers. OpenAI describes Astra as the first AI model with "critical" hacking abilities.
What Astra does
Zero-day vulnerabilities are software flaws that nobody knows about yet, which makes them especially dangerous. Astra doesn't just spot one flaw. It can string several together, turning a series of separate weaknesses into a single, working exploit. That's a level of automation that hasn't been seen before in an AI model, according to the company.
The process doesn't require a human to walk it through each step. Astra figures out the chain on its own, which is a big leap from earlier tools that needed a person to point them at a target and guide the attack.
Limited access for now
OpenAI is keeping Astra on a short leash. Only a small group of testers can use it, and the company hasn't said who those testers are or what criteria they had to meet. The tight access suggests OpenAI is aware of the risks. A model that can autonomously build exploits is a powerful tool, and the company appears to be moving carefully before deciding how — or whether — to release it more widely.
There's no word on when the testing phase might end, or what the next step would be. OpenAI hasn't announced a public release date, and it hasn't said whether Astra will ever be offered as a product or service.
Why "critical" matters
The "critical" rating is notable because it's the first time an AI model has been described that way. In cybersecurity, "critical" usually refers to the severity of a vulnerability — something that can be exploited remotely, without authentication, and with serious impact. Applying that label to an AI model's hacking ability signals that OpenAI considers Astra's capability to be on par with the most dangerous threats.
That raises questions about how the model will be kept out of the wrong hands. OpenAI hasn't detailed the safeguards it has in place beyond the limited tester group. The company also hasn't said whether it plans to share Astra's findings with software vendors so they can patch the flaws before they're exploited.
OpenAI hasn't said when Astra will be released, or whether it will ever be made broadly available. The testing phase is ongoing, and the company hasn't shared a timeline for when that might change. For now, the model exists in a controlled environment, and the rest of the world is waiting to see what happens when the testers start using it.
