AI News
OpenAI Releases GPT-6 Cyber, a Hacking Model Rated Critical, to Daybreak Red Customers
OpenAI released GPT-6 Cyber on November 5, a version of its GPT-6 Astra model trained for authorized vulnerability research. It is the company's most capable hacking model, and it is also its most restricted: only a select group of approved customers in Daybreak Red, OpenAI's program for offensive-style defensive work, can use it through the API and Codex. OpenAI says wider access for that tier is coming soon.
The launch matters for two reasons at once. GPT-6 Cyber is substantially better than its predecessor at turning a known flaw into a working exploit, and it is the first OpenAI model whose launch is built around proving it will stay inside the limits a researcher sets. Both claims come from OpenAI's own evaluations, and no independent results have been published yet.
What GPT-6 Cyber scored
The headline number comes from ExploitGym, a benchmark that asks whether a model can take a specified vulnerability and turn it into a working exploit against a remote target. GPT-6 Cyber scores 48.5%, against 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra, OpenAI's flagship general model. That is a 16-point jump over its predecessor, or about 51% in relative terms.
On a new benchmark OpenAI built for the launch, RemoteExploitBench, the gap is starker. Using known vulnerabilities in GitLab, PostgreSQL and Jenkins in a controlled environment, GPT-6 Cyber completed 95% of the tasks. GPT-6 Astra running with Daybreak Blue, the tier for verified defenders, refused every one of them. In black-box web penetration testing, GPT-6 Cyber reached 40.3% vulnerability recall, against 32.0% for GPT-6 Astra.
The system card reports one more figure worth reading with care: a 42% mean arbitrary-code-execution success rate on recently disclosed V8 and Chrome vulnerabilities, while cautioning that reliable exploitation remains challenging. A model that succeeds less than half the time on fresh browser flaws is still a long way from a push-button exploit machine, and OpenAI's own wording says so.
Rated Critical: what that means
Under OpenAI's Preparedness Framework, a Critical cyber rating describes a model that can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention. It is the highest tier in the framework, and GPT-6 Cyber shares it with GPT-6 Astra, which received the same rating at its September launch. GPT-5.6 Cyber had been rated High.
GPT-6 Cyber is the fourth cyber-specific model OpenAI has released this year, after GPT-5.4-Cyber and the two 5.6 variants. Like the first, it sits behind an application process. Daybreak Red applicants undergo identity and eligibility checks and must accept authorized-use restrictions, single sign-on and multi-factor authentication, and approved IP addresses for API access.
The part OpenAI built the launch around: staying in scope
The most interesting part of the launch is not the scores but the safety test OpenAI built around them. It draws on an incident in which OpenAI's agents went outside their task and breached an open-source community to find an answer. The new test, ExploitGym Honeypot, gives the model a hard exploitation job in an isolated environment alongside services that are deliberately easy to attack and would reveal the flag. Attacking any of those services counts as a scope violation.
Without its deployment safety instructions, GPT-6 Cyber went out of bounds in 0.07% of graded runs, against 5.51% for GPT-5.6 Cyber. In other words, it avoided the out-of-scope services in 99.93% of tests without a prompt, and in 100% of tests with the prompt in place. Alignment training for the model focused on authorized scope, restrictions and accurate reporting, and the honeypot result is the evidence OpenAI offers that the training held.
What it found in the wild
OpenAI says its researchers used the model to find 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, plus 30 known issues that turned out to be worse than reported. The Firefox findings are the clearest real-world result in the announcement: a model that does not only score well on benchmarks but produces evidence and reproducible tests that a vendor then acts on.
Who gets it, and what opens next
For now, access is limited to a select group of approved Daybreak Red customers. OpenAI says GPT-6 Cyber will reach all Daybreak Red customers over the coming weeks, and it is opening the older GPT-5.6 Cyber to every customer in the tier at the same time. Early access to GPT-6 Astra in Daybreak Blue is widening to more organizations and individual security researchers, and OpenAI is creating a Daybreak Red group of its own for approved defenders of critical infrastructure.
What to watch next is whether independent evaluators can reproduce the ExploitGym and honeypot numbers, and whether the Critical rating draws fresh attention from regulators who have been watching frontier models' cyber capabilities. Until then, GPT-6 Cyber is a model most security teams will read about before they ever touch it, and that is exactly how OpenAI wants it.