OpenAI Launches GPT-6 Cyber, a Security-Focused AI Model With Stronger Guardrails

OpenAI has introduced a new artificial intelligence model aimed at cybersecurity work, designed to help researchers identify software vulnerabilities, build working exploits and close security gaps while reducing the risk of misuse. The model, called GPT-6 Cyber, was announced on November 5 and is being positioned as both more capable and more tightly controlled than its predecessors.

According to OpenAI, the new model outperforms both GPT-5.6 Cyber and GPT-6 Astra across several security-focused tests. The company says it also follows instructions more reliably, which cuts down the chance that it acts on systems outside an authorized task. In testing, GPT-6 Cyber stepped outside its assigned boundaries in just 0.07 percent of cases, compared with 5.51 percent for GPT-5.6 Cyber.

On ExploitGym, a benchmark that measures whether an AI model can successfully exploit known software flaws, GPT-6 Cyber scored 48.5 percent. That is a 51 percent relative gain over GPT-5.6 Cyber, which scored 32.1 percent, and also ahead of GPT-6 Astra at 42.4 percent. In a separate web penetration testing evaluation, the model identified 40.3 percent of vulnerabilities, against 32 percent for GPT-6 Astra.

The model also completed 95 percent of tasks on RemoteExploitBench, which tests exploitation of vulnerabilities in controlled environments involving software such as GitLab, PostgreSQL and Jenkins. By comparison, GPT-6 Astra, running under Daybreak Blue restrictions, refused all tasks in that evaluation. GPT-6 Cyber is initially available to selected customers through OpenAI's Daybreak Red cybersecurity program.

Under OpenAI's Preparedness Framework, GPT-6 Cyber has been classified as Critical in cybersecurity, up from the High rating given to GPT-5.6 Cyber. The Critical label indicates that the model can independently identify and develop working zero-day exploits across many hardened, real-world critical systems. OpenAI highlighted that a major focus of the new model is preventing it from targeting systems beyond its assigned task.

The launch reflects growing attention on how powerful AI tools are governed, particularly when they can be used to probe and exploit software. For security teams, the pitch is a faster way to surface weaknesses before attackers find them, paired with tighter controls intended to keep the technology within authorized boundaries.