OpenAI launches GPT-6 Cyber, a Critical-rated security model
The Daybreak Red model hits 48.5% on ExploitGym and stays in scope 99.93% of the time, OpenAI says — but access stays gated.
OpenAI on Wednesday (Nov. 5) launched GPT-6 Cyber, a cybersecurity model it says is built to find and validate exploitable software flaws while staying inside the scope researchers set.
The company classifies the model as Critical for cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. That is the same Critical designation it gives mainline GPT-6 Astra. Access starts on the API and in Codex for a select group of Daybreak Red customers, with a wider Daybreak Red rollout still ahead.
On ExploitGym, which tests whether a model can exploit a specified vulnerability in a remote target, OpenAI reports GPT-6 Cyber at 48.5 percent, up from 32.1 percent for GPT-5.6 Cyber — about 16 percentage points, or a roughly 51 percent relative gain. The system card puts the same score at 48.45 percent and lists GPT-6 Astra at 42.43 percent.
"With the 6 Cyber model, we were able to make huge improvements on this evaluation," Fouad Matin, head of OpenAI's cyber preparedness and safety team, told reporters, according to The Verge. "It's over 99 percent of the time doing what you ask."
Staying in scope, and who gets it
One risk with cyber agents is that they wander into services that were never part of the job. OpenAI's ExploitGym Honeypot test, shaped by its earlier Hugging Face incident, gives the model easy-to-abuse helper services that sit outside the assigned target.
Without a deployment safety prompt, OpenAI says GPT-6 Cyber avoided exploiting those out-of-scope services in 99.93 percent of tests, up from 94.49 percent for GPT-5.6 Cyber. With the prompt, the company says the new model stayed in bounds 100 percent of the time.
Matin also told The Verge that evaluation awareness — when a model notices it is being tested and changes behavior — remains "super top of mind." OpenAI says it is pushing more production-like tests on real codebases to reduce that risk.
GPT-6 Cyber also leads OpenAI's new web penetration-testing recall eval at 40.3 percent, versus 32.0 percent for GPT-6 Astra, according to the announcement. On RemoteExploitBench, which uses known flaws in apps such as GitLab, PostgreSQL and Jenkins in a controlled lab, OpenAI says Astra running with Daybreak Blue refused every task, while GPT-6 Cyber completed 95 percent.
Daybreak is the gated program around that work. OpenAI frames Daybreak Blue as the track for verified defenders who need to validate bugs and fixes with fewer refusals, and Daybreak Red as the track for deeper exploit development and authorized testing with GPT-6 Cyber. Mainline Astra remains the model OpenAI pitches for code review, triage and patch writing.
The company says it will expand GPT-6 Cyber to all Daybreak Red customers soon and is also opening GPT-5.6 Cyber to that full group. Daybreak Blue alpha access for Astra is widening too. OpenAI has not published a consumer ChatGPT release or a public API price for GPT-6 Cyber.
Firefox, Linux and the defender pitch
OpenAI says the model is already in use on large, long-running security jobs. In its own research writeup, the company credits GPT-6 Cyber with developing 948 Linux privilege-escalation exploits — more than five times the 166 it attributes to GPT-5.6 Sol on the same harness — after reproducing 1,724 vulnerabilities. It says it disclosed 304 critical Linux findings to maintainers with candidate patches.
Separately, OpenAI says a campaign produced evidence for 81 previously unknown high-severity Firefox vulnerabilities, plus 30 known issues that looked worse than earlier reports. Mozilla has begun shipping fixes, OpenAI says.
"GPT-6 Cyber is a world-class model for cyberdefense," Brian Grinstead, a distinguished engineer at Firefox, said in a statement published by OpenAI. "Even after further hardening with multiple generations of frontier models, GPT-6 Cyber uncovered previously unknown high-severity vulnerabilities … and helped us ship fixes to millions of users."
Security teams named in OpenAI's post include Cisco, Calif, CodeRabbit, Palo Alto Networks, CrowdStrike, SentinelOne, Perplexity and Cognition. Dragos ran controlled OT lab evaluations that OpenAI says helped tune safeguards aimed at critical infrastructure.
Those safeguards, in OpenAI's telling, include cyber-specific monitoring, default IP allowlisting for Daybreak Red API access, Private Safety Processing for many zero-retention API setups, and new blocks on high-risk help against industrial and safety-critical systems. The company also points to a $1 billion Daybreak for Frontline Defenders grant program for resource-constrained teams.
None of the headline capability figures have been independently audited in public benchmarks outside OpenAI's materials. The Verge noted the Critical label means, in OpenAI's framework language, a model that can identify and develop functional zero-day exploits across many hardened real-world systems without human intervention — a claim that sits entirely on the vendor's evals and access rules.
OpenAI posted the launch on X at about 12:09 p.m. ET on Nov. 5 and is promoting a "Cyber at the Frontier" livestream on Nov. 12 at 10 a.m. PT (1:00 p.m. ET) for more cyber announcements. Until Daybreak Red opens more widely, GPT-6 Cyber remains a tool for approved defensive teams, not a general ChatGPT upgrade.
