OpenAI introduced GPT-6 Cyber on November 5, a specialized model for researching vulnerabilities, developing exploits, and validating flaws in authorized environments. The company classified it at the Critical cybersecurity level under its Preparedness Framework, up from the High level assigned to GPT-5.6 Cyber. This is not a general ChatGPT release: initial access is limited to approved Daybreak Red customers through the API and Codex, with gradual expansion planned within that same program.[1]

The clearest jump appears on ExploitGym, a test of whether a model can exploit a specified vulnerability in a remote target. GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber—a 16.4 percentage-point difference, or roughly a 51% relative improvement. On a new web penetration-testing evaluation conducted without source-code access, OpenAI reported 40.3% vulnerability recall, versus 32.0% for GPT-6 Astra.[1]

The specialization also changes how the model handles authorized offensive requests. On RemoteExploitBench, GPT-6 Astra operating through Daybreak Blue refused every task, while GPT-6 Cyber completed 95% of them in a controlled environment containing known vulnerabilities. That result does not imply universal effectiveness. The system card says that on FrontierCyber, GPT-6 Cyber solved 90 of 288 challenges, below GPT-6 Astra’s 96. Neither model solved any of the 15 highest-difficulty challenges, and no successful attacks were observed against fully hardened targets.[2]

OpenAI says the model has already been used to develop 948 Linux privilege-escalation exploits and identify 81 previously unknown high-severity Firefox vulnerabilities. These are company-reported results, not an independent audit, and should be read with that attribution. Even so, they indicate the scale a specialized agent can bring to reviewing large, security-critical codebases.[1]

The most sensitive issue is scope control. In a separate evaluation, GPT-6 Cyber avoided exploiting services outside the authorized target in 99.93% of tests without the deployment prompt and in 100% with that prompt. The system card describes additional layers including verified access, monitoring, Daybreak program restrictions, and training to respect boundaries. Those percentages come from scenarios designed by OpenAI; they do not establish that the model cannot cause harm or eliminate misuse risk.[1][2]

Independent coverage from TechCrunch places the launch within OpenAI’s broader push to give authorized researchers offensive tools and notes that Daybreak Red is the program’s most restricted access tier. The outlet also links the release to OpenAI’s separately announced $1 billion initiative for defenders working on critical infrastructure. That funding effort is distinct from the specific availability of GPT-6 Cyber.[3]

The relevant conclusion is not that AI has “solved” cybersecurity. Specialized offensive capability has crossed a threshold that its own developer considers critical, while access remains concentrated among a small, vetted group. The defensive value could be substantial if the model helps researchers find and patch flaws before attackers exploit them. The decisive test will be whether identity, scope, and monitoring controls continue to hold as the number of users, targets, and requests to deploy the model outside controlled laboratories increases.

*Illustration generated with AI for AI News; it does not depict an actual OpenAI interface.*

Sources

[1] Read More — GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents — OpenAI
[2] Read More — GPT-6 Cyber System Card — OpenAI Deployment Safety Hub
[3] Read More — OpenAI released a new cybersecurity model. Is that enough? — TechCrunch