OpenAI has rolled out Astra—a brand-new AI model that's gearing up for launch. For the first time, the company has given one of its systems a "critical" cybersecurity capability rating. According to OpenAI, Astra can independently hunt for vulnerabilities, craft exploits, and run complex attacks against protected systems.

What Astra Can Do

This model automates the entire offensive cyber playbook: everything from finding weak spots to building and executing multi-step attack chains. In test runs, Astra pulled off tasks that usually take a whole team of tools and security pros working together.

Testing Results

During trials, Astra uncovered two previously unknown zero-day vulnerabilities and chained them together in a series of exploits. The model managed to break into a hardened browser, escape the sandbox, and grab root access to the OS. OpenAI says Astra blows GPT-5.6 Sol out of the water when it comes to finding bugs and building exploits; in one test, it even scored a perfect 100%.

Launch and Access

Because of its powerful capabilities, OpenAI paused some of Astra's training earlier to beef up security measures. After launch, the most advanced cyber features will only be available to a select group of testers at first.