OpenAI Starts Rolling Out Astra, a Flagship AI Model Highly Skilled at Exploiting Security Flaws

OpenAI has started rolling out its newest frontier model, GPT‑6 Astra, OpenAI’s first system to reach a “Critical” cybersecurity rating. The company claims Astra can discover new security flaws and create working exploits in secure systems without needing very detailed instructions from humans. This also means that the model can create new ways to cause serious harm, not just increase existing risks—a major reason why OpenAI postponed and reevaluated parts of Astra’s launch.

Now, as of Thursday, Astra is on its way into the world. Access is currently restricted to trusted partners and members of the Daybreak cybersecurity coalition, OpenAI’s own cyber defense network. The company will soon allow more people to use it via ChatGPT Plus, Pro, Business, and Enterprise, as well as API and AWS users.

OpenAI reports that Astra achieved a perfect score on an exploit benchmark and located two zero-day vulnerabilities during internal testing. To reduce misuse, Astra is said to use stricter refusals on disallowed cyber requests and monitor production to halt tasks if unauthorized actions are detected. OpenAI also says that all this may sometimes slow or interrupt legitimate security work.