OpenAI has announced that its new GPT-6 Astra model has crossed the "Critical" cybersecurity capability threshold under the company's internal Preparedness Framework, making it the first model in OpenAI's lineup to receive that rating.

Alongside Astra, the company maintains a specialized sibling model called GPT-6 Cyber, which OpenAI says is already serving security teams at Mozilla, Palo Alto Networks, and Cisco.

What Astra Can Actually Do

According to OpenAI's research findings, GPT-6 Astra is capable of autonomously finding and exploiting zero-day vulnerabilities — flaws unknown to a program's maintainers until they are discovered, leaving no patch ready — across complex systems. The model can also construct full cyberattack plans starting from high-level goals.

On the ExploitBench evaluation, Astra posted a perfect score of 100%, a marked improvement over its predecessor, GPT-5.6 Sol, which scored 78.5% on the same test.

OpenAI's internal testing ran from September 1–4, 2026. During that window, Astra surfaced two previously unknown zero-day vulnerabilities. According to the findings, those two flaws will be disclosed to the maintainers of the affected software.

GPT-6 Cyber Goes to Work

OpenAI released GPT-6 Cyber on November 5, 2026. The model is a specialized variant built for controlled vulnerability research and exploit validation, with access limited to vetted customers only.

On the ExploitGym benchmark, GPT-6 Cyber scored 48.5%, up from the 32.1% recorded by the previous version, GPT-5.6 Cyber.

The research team reported finding 81 previously unknown high-severity vulnerabilities in Firefox. The team also said it developed working exploits for 948 Linux kernel privilege escalations — flaws that would let an attacker gain access levels beyond what the system permits — 588 of which were new. The Linux kernel underpins a large share of the servers, cloud platforms, and embedded devices in use today, which gives kernel-level findings reach well beyond a single product.

Mozilla credited OpenAI in its Firefox 159 security advisory for the vulnerabilities that were found.

Lee Klarich, Chief Product Officer and Chief Technology Officer of Palo Alto Networks, praised the model's effectiveness for defensive work. Mozilla engineer Brian Grinstead offered similar praise.

Why the Critical Rating Matters

OpenAI's Preparedness Framework is the company's internal system for grading how dangerous a model's capabilities could be in sensitive domains, with cybersecurity among those areas. Reaching "Critical" means OpenAI itself judges Astra's abilities to sit at the top of its risk scale.

OpenAI's answer, at least for the Cyber variant, appears to be gatekeeping. By restricting access to vetted customers and routing discoveries through responsible disclosure — the practice of reporting flaws privately to maintainers so fixes can be prepared before findings are made public — the company is attempting to preserve the defensive upside of the technology while limiting its offensive potential. The next observable step is the disclosure of the two zero-days Astra surfaced during September testing, which OpenAI says will go to the maintainers of the affected software, a pattern the Firefox 159 advisory credit already reflects.