Cyberweapon Out of the Box: OpenAI Deems Astra Model a Critical Threat for Finding 0-Days

Autonomous AI agents have definitively moved into the dual-use weapon category. On September 2, 2026, OpenAI made an unprecedented announcement: the upcoming `Astra` model has reached the "Critical" level under its cybersecurity Preparedness Framework.

During internal audits, the neural network demonstrated the ability not just to write code, but to autonomously discover previously unknown zero-day vulnerabilities, building highly complex exploitation chains. As a result, OpenAI is forced to urgently neuter the model, blocking access to its most powerful features for the general public. For the B2B market and macroeconomics, this is a top-level red alert. Yesterday’s fears from insurers and governments have been justified: an algorithm capable of autonomously hacking corporate infrastructures is entering the market. Enterprise spending on cybersecurity (Zero Trust systems) will skyrocket, as traditional firewalls are powerless against cognitive hacking.

Source: OpenAI / Reuters