Analysis
OpenAI's GPT-6 Cyber scores 48.5% on ExploitGym, up 51% on GPT-5.6 Cyber, and is limited to approved Daybreak Red customers
OpenAI says GPT-6 Cyber is 51% better on ExploitGym, strays out of scope under 1% of the time and found 81 Firefox flaws; only Daybreak Red customers get it.
On November 5, 2026, OpenAI released GPT-6 Cyber, a model tuned to find and exploit software vulnerabilities for authorized security work. OpenAI reports 48.5% on its ExploitGym benchmark against 32.1% for GPT-5.6 Cyber, which is a 51% relative gain, and says the model left out-of-scope services alone in 99.93% of a test without safety prompts. It says researchers used it to find 81 previously unknown high-severity Firefox flaws. For now only approved Daybreak Red customers can use it. Every figure here comes from OpenAI and has not been independently verified.
The essentials
- ExploitGym: 48.5% against 32.1% for GPT-5.6 Cyber. The 51% is relative; the gap is about 16 points.
- Staying in scope: in an internal honeypot test, GPT-6 Cyber avoided out-of-scope services 99.93% of the time without deployment prompts, against 94.49% for GPT-5.6 Cyber (roughly 5% stray).
- Real-world finds: 81 new high-severity Firefox flaws, which Mozilla rated severe and has started fixing.
- Access: select Daybreak Red customers, through the API and Codex, with wider access to Daybreak Red customers promised.
- Rating: OpenAI classifies it as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber.
What are the benchmark numbers, and what do they mean?
ExploitGym tests whether a model can exploit a given vulnerability in a remote target. GPT-6 Cyber scores 48.5%, GPT-6 Astra (the Daybreak Blue model) 42.4% and GPT-5.6 Cyber 32.1% (The Decoder). The Register describes ExploitGym as an OpenAI training evaluation, so the benchmark and the scoring are OpenAI's. In a separate web penetration test without source code, GPT-6 Cyber found 40.3% of the vulnerabilities against 32.0% for GPT-6 Astra (ShiftDelete.net).
How reliable is the "stays in scope" claim?
In the ExploitGym Honeypot test, GPT-6 Cyber avoided attacking out-of-scope services 99.93% of the time without deployment safety prompts and every time with them, while GPT-5.6 Cyber managed 94.49% (The Next Web). That is a controlled internal test with decoy services, not a measure of behavior in the field. OpenAI also requires IP allowlisting for Daybreak Red API access by default.
What did it find in Firefox and Linux?
OpenAI says researchers used it to find 81 previously unknown high-severity Firefox flaws, with reproducible proofs, and Mozilla has started landing fixes (Windows Report). In the Linux kernel, it reproduced 1,724 vulnerabilities and helped build working exploits for 948, of which 588 were previously unknown; these are reported with potential patches for maintainers (The Register).
Who can use it?
Daybreak Red is OpenAI's tier for authorized offensive security work such as exploit development and penetration testing, including against remote targets; Daybreak Blue is for validating vulnerabilities and fixes. GPT-6 Cyber is available to a select group of approved Daybreak Red customers and OpenAI says it will expand access over time.
Summary table
| Point | What OpenAI reports (as relayed by the press) | Source |
|---|---|---|
| ExploitGym | 48.5% vs 32.1% for GPT-5.6 Cyber (+51% relative) | The Decoder, Windows Report |
| Out-of-scope behavior | 99.93% in scope without prompts; 94.49% for GPT-5.6 Cyber | The Next Web |
| Firefox | 81 new high-severity flaws; fixes started | Windows Report |
| Linux kernel | 948 exploits, 588 previously unknown | The Register |
| Risk rating | Critical (GPT-5.6 Cyber: High) | ShiftDelete.net |
| Access | Select Daybreak Red customers, API and Codex | The Register |
Frequently asked questions
Is the 51% improvement 51 points?
No. It is a relative gain: 48.5% against 32.1% is about 16 points, or 51% more.
Is it available to the public?
No. Only approved Daybreak Red customers have access for now.
Method and limits
Sources as of November 5, 2026. We did not read OpenAI's announcement directly and rely on The Register, The Next Web, The Decoder, Windows Report, ShiftDelete.net and Finimize. All performance figures are OpenAI's own and unaudited. We tested nothing.
Sources
- The Register: OpenAI releases GPT-6 Cyber, and promises this one stays within 'rules of engagement' (verified 2026-11-05)
- The Next Web: OpenAI's GPT-6 Cyber found 81 new Firefox flaws and is rated 'Critical' (verified 2026-11-05)
- The Decoder: OpenAI's GPT-6 Cyber can both break into systems and propose how to fix them (verified 2026-11-05)
- Windows Report: OpenAI launches GPT-6 Cyber, access limited to approved Daybreak Red customers (verified 2026-11-05)
- ShiftDelete.net: GPT-6 Cyber, OpenAI's Critical-rated security model (verified 2026-11-05)
Spotted an error in this article? Report it · Our corrections policy
