OpenAI has officially unveiled GPT-6 Cyber, a powerful new specialized security model that the company reports is significantly more capable than its predecessor while exhibiting vastly improved discipline regarding the strict operational limits established by human researchers. The introduction of this advanced artificial intelligence model marks a notable step forward in the intersection of generative AI and cybersecurity, blending immense offensive potential with robust alignment measures designed to prevent misuse.
According to benchmark evaluations shared by the artificial intelligence developer, GPT-6 Cyber demonstrated a dramatic leap in performance on ExploitGym, a standardized platform that rigorously tests whether an AI system can successfully identify and exploit a given flaw within a remote target. On this benchmark, the new model achieved a success rate of 48.5%, representing a substantial increase compared to the 32.1% score recorded by its predecessor, GPT-5.6 Cyber.
Reflecting these heightened capabilities, OpenAI has formally rated GPT-6 Cyber as Critical for cybersecurity under its internal Preparedness Framework. This elevated classification places the model above the High designation that was previously assigned to GPT-5.6 Cyber, signaling both the profound power of the system and the heightened level of scrutiny required for its deployment and governance.
Alongside these raw capability metrics, OpenAI highlighted notable advancements in the model’s behavioral alignment and adherence to authorized boundaries. In a specialized alignment test designed to evaluate restraint, the model was assigned a primary exploitation task alongside a series of supporting services that were intentionally configured to be easy to hack but were explicitly designated as off-limits.
When tested without deployment safety prompts, GPT-6 Cyber successfully refrained from targeting the unauthorized decoy services in 99.93% of cases. This performance represents a clear improvement over GPT-5.6 Cyber, which achieved a restraint rate of 94.49% under the same unprompted conditions. Furthermore, when the safety prompts were actively applied, GPT-6 Cyber completely avoided the decoy services in every single test, never once overstepping its authorized operational boundaries.
OpenAI researchers noted that this tangible improvement in alignment reflects dedicated, targeted training aimed at ensuring
the model stays strictly within an authorized scope, respects operational restrictions, and reports its analytical results with a high degree of accuracy. Maintaining this delicate balance between offensive capability and strict guardrails remains a central focus for safety teams as frontier models grow increasingly proficient at identifying complex software vulnerabilities.
To demonstrate the real-world utility and depth of the model, OpenAI researchers directed GPT-6 Cyber toward the complex environment of the Linux kernel. During these rigorous tests, the AI model successfully reproduced 1,724 distinct vulnerabilities and systematically built functional privilege escalation exploits for 948 of them. For comparison, when the older GPT-5.6 Sol model was evaluated using the exact same testing tooling and environment, it produced only 166 reproducible exploits.
Capitalizing on these extensive findings, OpenAI has actively reported 304 critical security issues to software maintainers, complete with proposed fixes to help patch the underlying code. A similar exercise was conducted on the Mozilla Firefox browser, where the model successfully documented 81 new flaws. Mozilla deemed each of these newly discovered flaws to be of high severity, noting that the model backed every finding with comprehensive evidence and reproducible tests. In response to these AI-driven discoveries, Mozilla has already begun the process of rolling out necessary patches to protect end users.
To manage the inherent risks associated with such potent technology, OpenAI has integrated a suite of new protections directly into the deployment architecture. These new defensive measures include advanced monitoring systems designed to continuously check whether the actions taken by the model align precisely with the specific actions a human user is authorized to perform. Additionally, the system incorporates hard safeguards intended to prevent OpenAI’s models from knowingly or unknowingly aiding any efforts to compromise industrial control systems and other critical infrastructure targets.
OpenAI refined these critical safeguards through collaborative feedback from Dragos, a prominent industrial cybersecurity firm. Experts at Dragos tested the Daybreak Red models within a controlled laboratory environment to evaluate how effectively the safeguards could mitigate real-world threats against operational technology.
Rob T. Lee, the chief artificial intelligence officer at Dragos, emphasized the shifting paradigm of modern cybersecurity during these evaluations. He observed that the skill floor is dropping significantly for the reconnaissance and vulnerability research phases that traditionally must take place before an attacker can launch a disruptive campaign against targeted networks.
At present, access to GPT-6 Cyber remains strictly limited to select, approved customers participating in Daybreak Red, which is the specialized segment of OpenAI’s broader Daybreak program focused on advanced exploit work and authorized security testing. These approved users are able to leverage the capabilities of GPT-6 Cyber through the API and Codex interfaces.
Looking ahead, OpenAI plans to gradually expand access to the model for all participants in Daybreak Red, a group that is also receiving access to GPT-5.6 Cyber. Furthermore, a dedicated Daybreak Red group specifically tailored for approved critical infrastructure defenders will soon permit sanctioned, highly controlled testing of critical systems to bolster national and industrial security postures against emerging digital threats.

