OpenAI introduced GPT-6 Cyber on November 5, a model built for security research. It is meant to find and validate exploitable vulnerabilities while staying inside the scope researchers set for it. That second half is the actual claim.
The hard number sits in the capabilities section: OpenAI classifies the model as Critical in cybersecurity under its own Preparedness Framework. GPT-5.6 Cyber was rated High.
The benchmarks
On ExploitGym, which tests whether a model can exploit a named vulnerability in a remote target, GPT-6 Cyber scores 48.5% against 32.1% for GPT-5.6 Cyber — about 16 percentage points, which OpenAI reports as a 51% relative improvement. On a new offensive web pentesting evaluation, where models probe applications without source code, GPT-6 Cyber recalls 40.3% of vulnerabilities and GPT-6 Astra 32.0%.
RemoteExploitBench, also new, is the more telling one: known vulnerabilities in GitLab, PostgreSQL and Jenkins, deployed in a controlled environment. GPT-6 Astra running with Daybreak Blue refused every task. GPT-6 Cyber completed 95% of them; the single failure was an exploit it could not get to work remotely.
For scope adherence, OpenAI reports a separate evaluation: without its deployment prompt, the model avoided attacking out-of-scope services in 99.93% of tests, and with the prompt in 100%.
What it has already found
The 948 privilege escalation exploits in Linux were covered here yesterday. The new figure is the second one: 81 previously unknown high-severity vulnerabilities in Firefox, in a codebase that had already been searched extensively with frontier models. The model also produced candidate patches, and Mozilla has shipped fixes for many of those Firefox issues. Security teams at Cisco, Calif and CodeRabbit are using it, according to OpenAI.
Three tiers, and only the top one gets the model
Access runs through the Daybreak program, in tiers. The mainline models such as GPT-6 Astra cover code review, triage and patch development. Daybreak Blue helps verified defenders confirm vulnerabilities and fixes with fewer unnecessary refusals. Daybreak Red allows exploit development and authorized testing, and that is the only tier where GPT-6 Cyber runs — initially for a limited number of customers.
The classification is the news, not the benchmark
A company that puts its own model in the top risk tier of its own framework is saying two things at once: the capability is here, and it only leaves the building through an access gate. OpenAI writes that the window in which frontier models can secure software is limited. Whether three tiers are enough is the question TechCrunch puts in a headline the same day, and does not answer. Those 81 Firefox vulnerabilities are the first data point on which side the arithmetic lands.