On Thursday, OpenAI announced the release of its latest security model, GPT-6 Cyber. This new iteration promises enhanced capabilities and a stronger adherence to the boundaries set by researchers compared to its predecessor.
Performance Enhancements and Testing
The GPT-6 Cyber model demonstrated its prowess on the ExploitGym platform, achieving a score of 48.5%. This marks a significant improvement over the 32.1% scored by the previous GPT-5.6 Cyber model. OpenAI has classified GPT-6 Cyber as Critical in its Preparedness Framework, surpassing the High rating of the earlier version.
During alignment tests, GPT-6 Cyber showcased its discipline by refraining from exploiting intentionally vulnerable services 99.93% of the time without safety prompts, compared to 94.49% for GPT-5.6 Cyber. When safety prompts were utilized, the model completely avoided targeting these decoy services, underscoring its improved alignment training.
Key Findings and Security Contributions
OpenAI’s researchers conducted tests on the Linux kernel, where GPT-6 Cyber identified 1,724 vulnerabilities, developing functional exploits for 948 of these. In contrast, the earlier GPT-5.6 Sol model managed to identify only 166 vulnerabilities in similar tests. Additionally, OpenAI has reported 304 critical security issues to software maintainers, offering potential fixes for these vulnerabilities.
In another significant achievement, the model documented 81 high-severity flaws in Firefox, all backed by evidence and reproducible tests. Mozilla is currently working on deploying patches to address these issues.
Safeguards and Future Access
New protective measures have been integrated into the model to ensure its actions align with user permissions. Safeguards are also in place to prevent the model from assisting in the compromise of industrial controls and other critical systems. These measures were refined with feedback from cybersecurity firm Dragos, following controlled lab tests.
Currently, access to GPT-6 Cyber is exclusive to selected customers of OpenAI’s Daybreak Red program, which focuses on advanced exploit work. These customers can interact with the model through the API and Codex. OpenAI plans a gradual rollout to all Daybreak Red clients, including those using the GPT-5.6 Cyber. A dedicated group within Daybreak Red will facilitate authorized testing of critical infrastructure systems.
As OpenAI continues to advance its AI models, the introduction of GPT-6 Cyber represents a significant step forward in cybersecurity, providing enhanced tools for vulnerability research and defense.
