By Hamza Faisal
OpenAI has released GPT-6 Cyber, a specialised model built to find software flaws, turn them into working exploits and write the fixes — and it is keeping the tool out of public hands. The model, announced on November 5, is available only to a select group of approved customers in Daybreak Red, OpenAI's gated access programme for authorised security work.
GPT-6 Cyber is a cyber-specialised variant of GPT-6 Astra, OpenAI's flagship general model, with extra training for security investigations. OpenAI says it is designed for defensive research: penetration testing, red teaming and vulnerability discovery carried out with permission — not for open-ended hacking.
A sharper exploit engine — with a leash
The headline number is ExploitGym, a benchmark that asks a model to take a specified vulnerability and turn it into a working exploit against a remote target. GPT-6 Cyber scores 48.5 percent, up from 32.1 percent for its predecessor GPT-5.6 Cyber — a gain of about 16 percentage points, or roughly 51 percent in relative terms. It also edges out GPT-6 Astra itself, which scored 42.4 percent on the same test.
OpenAI now rates GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework, the top tier, up from High for GPT-5.6 Cyber and matching the rating given to Astra. The upgrade reflects a model materially better at offensive security research — and more dangerous in the wrong hands.
That is why the company has paired the capability with scope controls. The model is trained to stay inside the boundaries a researcher sets, and on a test built after an earlier incident in which OpenAI agents strayed outside their task, GPT-6 Cyber avoided out-of-scope services in 99.93 percent of cases without its deployment prompt — and 100 percent of the time with it. On a new benchmark, RemoteExploitBench, it completed 95 percent of tasks, while GPT-6 Astra running under the broader Daybreak Blue programme refused all of them.
What it found in the wild
OpenAI says its researchers used GPT-6 Cyber to uncover 81 previously unknown vulnerabilities in Firefox that Mozilla rated high severity, along with 30 known issues that turned out to be worse than first reported. Mozilla has shipped fixes for many of them. The company also says the model produced 304 critical findings for software maintainers, each with a candidate patch.
The release follows a demonstration published a day earlier, in which OpenAI's agents reproduced 1,724 Linux kernel vulnerabilities and built working exploits for 948 of them.
Vetted access only — and a premium price
Access starts with a small group of approved Daybreak Red customers through OpenAI's API and Codex, expanding to every customer in the tier over the coming weeks. Applicants face identity and eligibility checks, and must use single sign-on and multi-factor authentication; API requests must come from approved networks by default. Critical-infrastructure defenders will get a Daybreak Red group of their own. OpenAI is also opening the older GPT-5.6 Cyber to all Daybreak Red customers and widening Daybreak Blue access to GPT-6 Astra.
The new model also carries a premium price: OpenAI's developer documentation lists GPT-6 Cyber at $20 per million input tokens and $100 per million output tokens, with cached input at $2 per million — higher than the published rates of comparable models from Anthropic and Google.
There is no public access anywhere, and OpenAI has not said if that will change. For now, its most capable hacking model belongs to the defenders it has vetted, at a price that keeps it firmly in professional hands.
