OpenAI Ships GPT-6 Cyber: First AI Model Rated 'Critical' Under Its Own Safety Framework

OpenAI launched GPT-6 Cyber on November 6, 2026 — a model purpose-built for security research and the first AI system the company has ever rated "Critical" under its own Preparedness Framework. That classification sits one notch above the previous "High" ceiling, meaning OpenAI is publicly acknowledging it has shipped a model it considers potentially dangerous.

The launch is not theoretical. Access to GPT-6 Cyber is restricted to eligible customers — organizations with verified authorization for vulnerability research and exploit validation. The model is not available through standard ChatGPT or the general API tier.

What the "Critical" Rating Actually Means

OpenAI's Preparedness Framework is an internal risk evaluation system that grades AI models on potential for harm across five categories: cybersecurity, biological threats, radiological/nuclear threats, chemical threats, and autonomy. Prior to GPT-6 Cyber, no shipped OpenAI model had ever received a Critical rating in any category. The company has previously said it would not deploy a model scoring Critical unless it had mitigations in place that reduced residual risk to High or below.

The fact that GPT-6 Cyber carries a Critical rating and is being deployed — with restrictions — signals that OpenAI believes its access controls and use-case gating constitute adequate mitigation. That is a significant judgment call, and one that industry observers and AI safety researchers will examine closely.

The 948 Linux Vulnerabilities Precedent

GPT-6 Cyber did not arrive without evidence of capability. As previously reported by NexChron, a predecessor model autonomously identified and patched 948 Linux kernel vulnerabilities for $1.2 million in API costs — work that would have required dozens of security engineers and months of manual effort. That deployment was framed as defensive, but the same capability inverted is an offense-grade exploit generator.

GPT-6 Cyber is the production version of that capability class: trained specifically for security workflows, structured to interpret CVE databases, produce proof-of-concept code, and validate whether a patch fully closes a vulnerability. For security teams with the authorization to use it, that is a force multiplier. For anyone who finds a way around the access gates, it is a different kind of problem.

Who Gets Access

According to OpenAI, eligible customers include organizations conducting authorized penetration testing, vulnerability research programs, and exploit validation for defensive purposes. The model is not available to general enterprise customers or developers through standard API access. OpenAI has not disclosed the vetting criteria publicly, but the access model resembles restricted research agreements rather than normal commercial deployment.

Key launch details:

• Model name: GPT-6 Cyber
• Launch date: November 6, 2026
• Preparedness rating: Critical (first ever for a deployed OpenAI model)
• Access: Restricted to verified security research organizations
• Use cases: Authorized vulnerability research, exploit validation, defensive security

What This Means for the Security Industry

For authorized security teams, GPT-6 Cyber is the most significant AI tool ever made available to the profession. Penetration testers, red teams, and vulnerability researchers routinely spend weeks on work that an AI rated Critical under a major lab's own safety framework can automate. The economics of offensive security research shift dramatically.

For the broader software ecosystem, the launch raises an uncomfortable question: OpenAI controls the access gates today, but the capability now demonstrably exists. Other labs — some without OpenAI's public safety commitments — are watching. If GPT-6 Cyber-class capability leaks, is replicated by a less constrained model, or is accessed by a threat actor who falsifies authorization credentials, the consequences extend well beyond the security research community.

The NCSC warned this week that AI-powered attacks on critical infrastructure are likely by 2028. GPT-6 Cyber is evidence that the offensive capability timeline may be shorter than any government agency has publicly estimated.

What to Watch

Whether any unauthorized access attempts emerge against GPT-6 Cyber's access controls in the coming months will be an early test of the model's operational safety profile. Watch also for competing labs — particularly those without public Preparedness frameworks — to announce capability-equivalent security-focused models with fewer restrictions.