OpenAI has introduced GPT-6 Cyber, a new artificial intelligence model specifically designed for cybersecurity research and vulnerability discovery. The company classified the model as "Critical" under its Preparedness Framework, a designation indicating it can identify and develop functional zero-day exploits in hardened real-world systems without human intervention. This is an upgrade from the previous GPT-5.6 Cyber model, which held a "High" classification.
During testing, the model demonstrated a significant increase in reliability regarding task boundaries. OpenAI reported that GPT-6 Cyber stayed within its assigned scope 100% of the time when using deployment prompts, compared to a 99.93% success rate without them. This represents a notable improvement over GPT-5.6 Cyber, which went out of bounds in approximately 5% of tests. Fouad Matin, the head of OpenAI’s cyber preparedness and safety team, noted that the model is "over 99 percent of the time doing what you ask."
The model has already been utilized to identify 81 previously unknown high-severity vulnerabilities in Firefox, many of which Mozilla has since patched. Additionally, researchers used the model to develop working exploits for 948 Linux kernel vulnerabilities. Beyond open-source software, the model is being used by security teams at Cisco, Calif, and CodeRabbit for code review and vulnerability research, and OpenAI has employed it to secure its own internal infrastructure.
In performance benchmarks, GPT-6 Cyber outperformed both its predecessor and OpenAI's flagship GPT-6 Astra model. On the ExploitGym benchmark, which measures a model's ability to exploit specific vulnerabilities, GPT-6 Cyber scored 48.5%, surpassing GPT-6 Astra's 42.4% and GPT-5.6 Cyber's 32.1%. The model also achieved a 40.3% recall rate in offensive web penetration testing, compared to 32.0% for GPT-6 Astra.
To mitigate risks associated with autonomous agents, OpenAI implemented new safeguards to detect and disrupt misuse, specifically targeting attacks on critical infrastructure like water treatment facilities and power plants. The company also addressed concerns regarding "evaluation awareness," a phenomenon where models recognize they are being tested and alter their behavior. Matin stated that the company is prioritizing production-like evaluations involving real codebases to reduce this risk.
GPT-6 Cyber is currently being rolled out to select members of OpenAI’s Daybreak Red program, which is intended for organizations performing authorized defensive cybersecurity work. The company plans to expand access to all Daybreak Red customers in the coming weeks.
