OpenAI is preparing to roll out Astra, its first model to hit a “Critical” cyber threat tier. Here is what its autonomous hacking skills mean for security teams.
OpenAI has a powerful new model waiting in the wings, and it comes with serious hacking capabilities. Meet Astra.
This is the company’s first model to officially reach the “Critical” cybersecurity threshold under its internal Preparedness Framework.
Astra didn’t merely solve routine code puzzles during internal benchmarks. It independently uncovered unknown zero-day vulnerabilities, escaped sandbox environments, and built full exploit chains across hardened operating systems- all with minimal human guidance and surprisingly low compute power.
That last detail is what makes security experts sit up straight. When high-level offensive capabilities get cheaper to run, the digital landscape changes rapidly.
However, this isn’t a story about a rogue AI running wild. It’s a story about responsible scaling in real time. OpenAI briefly paused development in August to build tighter guardrails, boost its cyber jailbreak refusal rate to 91.5%, and design strict inference monitoring.
And now OpenAI is gating full offensive access exclusively for vetted security defenders through a tier called Daybreak Blue- rather than dropping these capabilities onto the open web. The model is also undergoing a formal U.S. government pre-release review.
This dual-use dynamic is undeniably tricky.
Security teams might fix every flaw Astra finds, but an attacker only needs a single open window of opportunity. But still, putting these tools directly into the hands of cyber defenders could be the first step in fundamentally tilting the balance toward a safer web.
OpenAI is setting a clear blueprint for how frontier AI risks should be managed.
