OpenAI launches GPT-6 Cyber, its first 'Critical' cyber model, for vetted red teams only
The model scores 48.5% on ExploitGym, found 81 high-severity Firefox bugs and, OpenAI says, stayed in scope in 99.93% of honeypot tests. Access is limited to approved Daybreak Red customers.
TL;DR
- OpenAI introduced GPT-6 Cyber on Nov. 5, 2026 and classified it as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber.
- GPT-6 Cyber scored 48.5% on ExploitGym, versus 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra, OpenAI and The Verge reported.
- OpenAI says the model found 81 previously unknown high-severity Firefox vulnerabilities and avoided out-of-scope exploitation in 99.93% of honeypot tests without a deployment prompt.
- Access starts with a select group of approved Daybreak Red customers via the API and Codex, with IP allowlisting required by default.
OpenAI has built a model it rates as dangerous enough to need a gatekeeper. On Nov. 5, 2026 the company introduced GPT-6 Cyber and classified it as Critical in cybersecurity under its Preparedness Framework, up from High for its predecessor GPT-5.6 Cyber. According to The Verge, that tier means a model "can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention."
The numbers
On ExploitGym, which tests whether a model can exploit a specified vulnerability in a remote target, GPT-6 Cyber scored 48.5%, up from 32.1% for GPT-5.6 Cyber, a 51% relative improvement. The Verge reports OpenAI's flagship GPT-6 Astra scored 42.4%. On a new web penetration-testing evaluation, GPT-6 Cyber reached 40.3% recall versus 32.0% for Astra. On RemoteExploitBench, Astra refused 100% of tasks; GPT-6 Cyber completed 95%, OpenAI says.
OpenAI says its researchers used the model to identify 81 previously unknown high-severity vulnerabilities in Firefox, and Mozilla has released fixes for many of them. It was also the model behind the 948 Linux privilege-escalation exploits OpenAI disclosed earlier, which The Register says include 588 previously unknown bugs. Security teams at Cisco, Calif and CodeRabbit are already using it.
Staying inside the lines
The bigger claim is about behavior. In an evaluation called ExploitGym Honeypot, modeled on the Hugging Face incident, the model gets easy-to-exploit side services that would reveal the answer. Without deployment safety prompts, GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of cases, up from 94.49% for GPT-5.6 Cyber; with the prompts, 100%. "It's over 99 percent of the time doing what you ask," Fouad Matin, head of OpenAI's cyber preparedness and safety team, told reporters, per The Verge. OpenAI declined to tell The Register what happened in the remaining 0.07% of cases.
Access is narrow. GPT-6 Cyber is available through the API and Codex to a select group of approved Daybreak Red customers, with broader Daybreak Red access to follow. OpenAI requires IP allowlisting for Daybreak Red API access by default and has added new checks to block high-risk activity against critical systems such as power plants and water treatment facilities.
What it means for crypto and Web3 teams – Analysis
Smart contracts, bridges and node software are exactly the kind of hardened, high-value code this model is built to break. If an AI can find 81 new high-severity bugs in Firefox, a codebase OpenAI describes as heavily tested, assume it can find bugs in your protocol too. The defensive question is no longer whether to use AI auditing, but whether you get there before someone with similar tools does. Gated access helps, but it is a policy, not a guarantee. This is not investment advice.
FAQ
What is GPT-6 Cyber?
A cybersecurity-tuned OpenAI model, introduced Nov. 5, 2026, that finds and validates exploitable vulnerabilities. OpenAI classifies it as Critical in cybersecurity under its Preparedness Framework.
How does it compare with earlier models?
On ExploitGym it scored 48.5%, versus 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra, according to OpenAI and The Verge.
Who can use GPT-6 Cyber?
Initially a select group of approved Daybreak Red customers through the API and Codex; OpenAI says it will expand access to all Daybreak Red customers.
Does it stay within authorized scope?
In OpenAI's ExploitGym Honeypot test, it avoided out-of-scope exploitation in 99.93% of cases without deployment prompts and 100% with them. OpenAI did not detail the remaining 0.07% to The Register.
Sources
News summary based on the sources above. Not investment, legal or tax advice. Editorial policy & corrections · Report an error
Building in Web3? Turn the news into trust.
Positioning, KOL programs, LinkedIn systems and claim-safe content for Web3 and fintech teams.
Related
OpenAI pledges to pay full power costs and drop NDAs with officials after voters reject data centers
Two days after voters backed data center restrictions in 40 of 41 local ballot measures, OpenAI published a National Community Compact and asked lawmakers to turn similar rules into law.
OpenAI buys Pierre, Meta buys Conductor: big labs absorb AI coding tools in one day
Both deals were announced on Nov. 4, 2026. Pierre's Code Storage will fold into Codex; Conductor stops sign-ups and deletes cloud data from Nov. 20, 2026.
OpenAI's cyber pipeline produced 948 Linux root exploits in one run for $1.2 million in API costs
In mid-October 2026 the company says agents reproduced 1,724 vulnerabilities and built working local privilege-escalation proofs for a stock Linux desktop — then generated patches for maintainers.