OpenAI has officially unveiled GPT-6 Cyber, a powerful new specialized security model that the artificial intelligence company says marks a dramatic leap forward in offensive and defensive cyber capabilities. According to the company, the new model is not only far more capable than its predecessor, GPT-5.6 Cyber, but also significantly more disciplined when it comes to respecting the strict operational boundaries and safety limits established by human researchers.
The announcement comes at a time of rapid evolution in the intersection of artificial intelligence and cybersecurity, where large language models are increasingly being evaluated for their potential to both discover critical software vulnerabilities and protect vital digital infrastructure. OpenAI’s introduction of GPT-6 Cyber highlights this dual-use reality, demonstrating unprecedented prowess in finding system flaws while simultaneously implementing advanced guardrails designed to prevent misuse.
In rigorous benchmark evaluations, GPT-6 Cyber showcased a massive improvement in technical performance. On ExploitGym, a specialized evaluation framework designed to test whether an artificial intelligence system can successfully exploit a given flaw in a remote target, the new model achieved an impressive 48.5% success rate. This represents a substantial jump from the 32.1% score recorded by GPT-5.6 Cyber, underscoring how quickly generational advancements are increasing the technical sophistication of automated exploit generation.
Reflecting these heightened capabilities, OpenAI has rated GPT-6 Cyber as "Critical" for cybersecurity under its internal Preparedness Framework. This classification places it a tier above the "High" designation that was previously given to its predecessor, GPT-5.6 Cyber, indicating that the model possesses capabilities that demand rigorous monitoring, specialized deployment restrictions, and advanced technical safeguards.
Despite its expanded capacity to interact with complex systems and execute sophisticated exploits, OpenAI emphasizes that GPT-6 Cyber exhibits a remarkable adherence to safety controls. In one detailed alignment test administered by researchers, the model was given a designated exploitation task alongside supporting services that were intentionally designed to be easy to hack, but were explicitly placed off-limits by the research team.
When tested without deployment safety prompts, GPT-6 Cyber refrained from exploiting the unauthorized decoy services in an astounding 99.93% of cases. This performance stands in sharp contrast to GPT-5.6 Cyber, which maintained a lower restraint rate of 94.49% under identical conditions. Furthermore, when the model was evaluated with safety prompts enabled, it never targeted the decoy services, showing a perfect adherence to its authorized operational scope.
OpenAI researchers noted that this tangible improvement in alignment is the direct result of dedicated, specialized training designed to help the model stay strictly within authorized boundaries, respect designated operational restrictions, and report its findings with high accuracy.
To demonstrate the real-world utility and raw power of the new model, OpenAI researchers pointed GPT-6 Cyber directly at the Linux kernel. In this demanding real-world test, the model successfully reproduced 1,724 known vulnerabilities and went on to construct functional privilege escalation exploits for 948 of them. For comparison, GPT-5.6 Sol, when tested using the exact same underlying tooling and parameters, produced only 166 functional exploits.
Following these extensive research findings, OpenAI has already taken proactive steps to remediate the discovered issues, reporting 304 critical security issues directly to the respective maintainers alongside proposed fixes to help secure the broader open-source ecosystem.
The model’s analytical capabilities were similarly demonstrated when applied to the Mozilla Firefox browser. During evaluations, GPT-6 Cyber documented 81 new flaws that Mozilla security teams independently deemed to be of high severity. The model backed each discovered vulnerability with comprehensive evidence and reproducible test cases, prompting Mozilla to begin rolling out patches to protect its user base.
To manage the inherent risks associated with deploying such a potent security model, OpenAI has introduced a suite of new protections alongside GPT-6 Cyber. These safeguards include advanced real-time monitoring systems that continually check whether the model’s active operations line up precisely with what an individual user is legally and operationally permitted to do. Additionally, new safeguards have been implemented specifically to stop OpenAI’s models from knowingly aiding any efforts to compromise industrial control systems and other critical infrastructure.
OpenAI refined these critical infrastructure safeguards by incorporating direct feedback from Dragos, a prominent industrial cybersecurity firm. Dragos tested early iteration models, known as Daybreak Red, within a controlled laboratory environment to evaluate how effectively AI tools could be integrated into the defense of vital industrial networks without introducing unacceptable risks.
Dragos Chief AI Officer Rob T. Lee commented on the shifting landscape of cybersecurity in light of these developments, noting that the skill floor is dropping significantly for the initial reconnaissance and vulnerability research that typically precedes a cyberattack. This shift places an even greater emphasis on the need for robust defense mechanisms that match or exceed the capabilities available to potential adversaries.
At present, access to GPT-6 Cyber is strictly limited to select approved customers participating in the Daybreak Red initiative, a specialized program managed by OpenAI focused on advanced exploit work and authorized security testing. Approved participants can access the model through dedicated application programming interfaces and coding environments like Codex.
Looking ahead, OpenAI plans to gradually expand access to the model for all Daybreak Red customers, who are also receiving access to GPT-5.6 Cyber as part of their program enrollment. Furthermore, a dedicated Daybreak Red group specifically tailored for approved critical infrastructure defenders will soon permit sanctioned, controlled testing of critical operational technology systems, helping organizations harden their defenses against increasingly automated threats.






