OpenAI unveiled GPT-6 Cyber on Thursday, a model dedicated to computer security that the company presents as significantly more performant than its predecessor. According to OpenAI, it also better respects the limits set by researchers.
- OpenAI’s GPT-6 Cyber outperforms GPT-5.6 Cyber on ExploitGym, scoring 48.5% against 32.1%.
- OpenAI rates the model «Critical» under its Preparedness Framework, one level above its predecessor.
- The model respects scope limits better and avoids out-of-scope decoy targets more often.
- Access stays restricted to approved Daybreak Red clients while OpenAI gradually widens it.
On ExploitGym, a test that checks if an AI can exploit a given vulnerability on a remote target, the new model achieves 48.5%, compared to 32.1% for GPT-5.6 Cyber. OpenAI classifies GPT-6 Cyber at the ‘Critical’ level for cybersecurity, according to its Preparedness Framework, its risk assessment framework. This level is above ‘High’, which was assigned to GPT-5.6 Cyber.
Behavior tested against decoy targets
In one of its alignment tests, the model was given an exploitation task accompanied by ancillary services, deliberately easy to hack but outside the authorized scope. Without security instructions at the time of deployment, it refrained from exploiting them in 99.93% of cases, compared to 94.49% for GPT-5.6 Cyber. With these instructions, it never targeted these decoy services.
‘This alignment improvement reflects dedicated training aimed at staying within the authorized scope, respecting restrictions, and reporting results accurately,’ states OpenAI.
Vulnerabilities identified in Linux and Firefox
OpenAI researchers applied the model to the Linux kernel, where it reproduced 1,724 vulnerabilities and built functional privilege escalation exploits for 948 of them. Tested with the same tools, GPT-5.6 Sol was also evaluated on this task. OpenAI indicates it has reported 304 critical issues to the project maintainers, accompanied by proposed fixes.
In the Firefox browser, the model documented 81 new vulnerabilities that Mozilla deemed high severity, each backed by evidence and reproducible tests. Mozilla has begun deploying the fixes.
Safeguards and restricted access
OpenAI is implementing new protections, including a monitoring system that checks if the model’s actions correspond to what the user is authorized to do. Other safeguards are meant to prevent its models from helping to compromise industrial control systems and other safety-critical systems. These measures were refined based on feedback from Dragos, a company specialized in industrial cybersecurity, which tested Daybreak Red models in a controlled lab.
‘The skill bar is lowered for the reconnaissance and vulnerability research needed before an attack,’ said Rob T. Lee, AI Director at Dragos.
For now, access is limited to certain approved Daybreak Red clients, the component of OpenAI’s Daybreak program dedicated to advanced exploit work and authorized testing. They can use GPT-6 Cyber via the API and Codex. OpenAI plans to gradually open the model to all Daybreak Red clients, who also receive GPT-5.6 Cyber. A Daybreak Red group dedicated to approved critical infrastructure defenders will additionally allow supervised testing on sensitive systems.
