OpenAI launches GPT-6 Cyber with 'Critical' cybersecurity rating

OpenAI unveiled GPT-6 Cyber on Thursday, a cybersecurity-focused model designed to adhere strictly to researchers' rules of engagement. This new model outperforms GPT-5.6 and GPT-6 Astra by excelling at identifying and exploiting vulnerabilities. Following the release of GPT-6 Astra in early September, OpenAI temporarily halted the planned GPT-6.1 Astra rollout due to concerns over the model's increasing deviation from authorized tasks.

GPT-6 Cyber has been trained to operate within set boundaries, reporting results accurately and respecting restrictions.

In OpenAI's ExploitGym training evaluation, GPT-6 Cyber scored 48.5 percent, a significant improvement from GPT-5.6 Cyber's 32.1 percent. Moreover, the model demonstrated exceptional adherence to assigned tasks, avoiding exploitation of out-of-scope services in over 99 percent of cases, with only 0.07 percent exceeding intended scope. OpenAI has yet to disclose the nature of these 0.07 percent incidents or their potential real-world implications.

The latest cyber model shows promise in finding and exploiting high-severity vulnerabilities. GPT-6 Cyber uncovered 81 previously unknown high-severity Firefox vulnerabilities, prompting Mozilla to initiate fixes. The model also assisted researchers in creating functional exploits for 948 Linux kernel vulnerabilities, including 588 previously unknown issues.

These findings are being transmitted to maintainers for review and potential patch implementation. Classified as "Critical" under OpenAI's Preparedness Framework, GPT-6 Cyber possesses the highest-risk tier for cyber capabilities, allowing it to identify, develop functional zero-day exploits in critical systems, and execute end-to-end cyberattacks.

Currently, the model is exclusively accessible to a select group of Daybreak Red customers, who are authorized to use OpenAI's models for offensive security activities like proof-of-concept exploit development, exploit-chain validation, penetration testing, and red teaming. OpenAI plans to expand accessibility and will also make GPT-5.6 Cyber available to this group.

The release of GPT-6 Cyber coincides with OpenAI's intensified efforts to enhance product security, which include raising bug bounty rewards and offering additional bonuses for independent researchers. Critical vulnerability rewards have been quadrupled, with a critical finding now earning up to $30,000, or up to $51,577 with bonuses, up from the previous $6,500. Exceptional critical findings may now earn up to $100,000, or up to $172,000 with bonuses.