OpenAI's Astra becomes its first AI model with 'Critical' hacking abilities

OpenAI's Astra has become the first of the company's AI models to demonstrate hacking abilities rated “Critical,” according to a Decrypt report. The rating marks a visible step up in what OpenAI's own models are measured to do in security-sensitive tasks.

The report does not say which evaluations Astra ran, but the Critical label points to strong performance in vulnerability exploitation, penetration and other offensive-security scenarios.

For the security industry, that kind of capability is a double-edged sword. A model with high-grade attack abilities could automate penetration testing, vulnerability discovery and red-team exercises, freeing specialists from repetitive work; but the same abilities, misused, could lower the barrier to real-world attacks.

More fundamentally, the rating moves model safety beyond refusing harmful requests and toward governing models that are themselves capable of attacking systems. For OpenAI, keeping Astra useful while containing that capability is a harder problem than alignment alone.

Watch next for how OpenAI draws the boundaries around Astra, whether it opens the capability to security researchers, and how regulators respond to frontier models with rated attack abilities.