OpenAI's GPT-6 Astra lands in its top-tier cyber risk bracket

OpenAI says GPT-6 Astra is the first model in its lineup to cross the Critical cybersecurity threshold under its Preparedness Framework, the company's internal system for grading how dangerous a model's capabilities could get in sensitive areas. Cybersecurity is one of those areas, and Critical sits at the top of the risk scale. In other words, OpenAI itself is judging Astra's abilities as dangerous as its own framework allows.

The findings are stark. Astra can autonomously find and exploit zero-day vulnerabilities across complex systems, and can build full cyberattack plans starting from high-level goals. On the ExploitBench evaluation it posted a perfect 100%, well clear of predecessor GPT-5.6 Sol at 78.5%. During internal testing from September 1-4, 2026, Astra surfaced two zero-day vulnerabilities that were previously unknown; both will be disclosed to the maintainers of the affected software.

Alongside Astra, OpenAI has a narrower tool: GPT-6 Cyber, released November 5, 2026, and built for controlled vulnerability research and exploit validation. Access is restricted to vetted customers only. On ExploitGym it scored 48.5%, against 32.1% for the previous GPT-5.6 Cyber. It is already serving security teams at Mozilla, Palo Alto Networks and Cisco.

The defensive haul is substantial. The research team reported 81 previously unknown high-severity vulnerabilities in Firefox, plus working exploits for 948 Linux kernel privilege escalations, 588 of which were new. Mozilla credited OpenAI in its Firefox 159 security advisory for the vulnerabilities found.

Palo Alto Networks CPO and CTO Lee Klarich praised the model's effectiveness for defensive work, and Mozilla engineer Brian Grinstead did the same. For the Cyber variant, OpenAI's answer is gatekeeping: limit access to vetted customers and push discoveries through responsible disclosure, a bid to keep the defensive upside while capping the offensive downside.