OpenAI presented on Thursday (5) the GPT-6 Cyber, an artificial intelligence model specialized in the discovery, authorized exploitation, and fixing of security vulnerabilities. The tool debuts with restricted access for approved customers of the Daybreak Red program and received a classification Critical in cybersecurity in the company's own risk assessment system.
In the ExploitGym benchmark, which assesses the ability to exploit specific vulnerabilities in controlled targets, the new model achieved a 48.5% success rate, compared with 32.1% for GPT-5.6 Cyber. The difference represents an advance of 16.4 points in percentage terms, equivalent to a relative improvement of approximately 51%.
The Critical classification, defined by OpenAI's Preparedness Framework program, reflects the level of capability demonstrated by the system in cybersecurity tasks. The previous model was classified at the High level.
What the tests show and who can access the model
In addition to performance in flaw exploitation, OpenAI assessed GPT-6 Cyber's ability to respect the limits established for each task. In the ExploitGym Honeypot test, the model avoided attacking services outside the authorized scope in 99.93% of cases without additional security instructions for deployment, a result that reached 100% with those instructions.
The percentages correspond to specific evaluations. They do not mean that the model can exploit any system with the same success rate or that it is completely protected against unauthorized behavior.
The company also disclosed results of research conducted on real projects. According to the statement, its researchers used GPT-6 Cyber to develop 948 functional privilege-escalation exploits in the Linux kernel and identify 81 previously unknown high-severity vulnerabilities in Firefox. Mozilla has already made fixes available for several of the reported flaws.
Unlike the benchmarks, these numbers represent vulnerability research results reported by the organization, not general performance rates of the model.
GPT-6 Cyber is initially available through the API and Codex, exclusively to a selected group of approved Daybreak Red customers. The program requires identity verification and additional security controls. OpenAI intends to gradually expand availability to the other approved participants, but has not announced a specific date for that expansion.
