OpenAI launched GPT-6 Cyber on Nov 5, 2026 — Critical under the Preparedness Framework, stronger on ExploitGym, gated to Daybreak Red. What the numbers say and who can use it.
OpenAI published GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents ↗ on November 5, 2026. The model is a cyber-specialized variant aimed at authorized vulnerability discovery, exploit development, and web pentesting — not a general chat upgrade.
The company’s own X account posted the same day:
“Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment… GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time.”
— @OpenAI ↗, Nov 5, 2026
The Verge ↗ independently covered the launch the same afternoon, quoting OpenAI’s Fouad Matin on staying in bounds “over 99 percent of the time.”
From OpenAI’s post and system card ↗:
| Metric | GPT-6 Cyber | Prior / comparison |
|---|---|---|
| ExploitGym (exploit a specified vuln) | 48.5% | GPT-5.6 Cyber 32.1% (~51% relative gain) |
| Out-of-scope honeypot avoidance (no deploy prompt) | 99.93% | GPT-5.6 Cyber 94.49% |
| Same, with deploy prompt | 100% | — |
| Web pentest vuln recall | 40.3% | GPT-6 Astra 32.0% |
| Preparedness Framework cyber level | Critical | GPT-5.6 Cyber was High |
OpenAI also says researchers used GPT-6 Cyber to help develop 948 Linux privilege-escalation exploits and identify 81 previously unknown high-severity Firefox issues (Mozilla began landing fixes after disclosure). Those are OpenAI’s disclosed campaign results — treat them as the lab’s claim, not an independent audit.
Access is gated: Daybreak Red customers only at first (API + Codex), with expansion planned. Daybreak Blue remains the defensive tier with fewer refusals on mainline models; Red is the deeper offensive / authorized-testing lane. GPT-6 Cyber’s model ID in the API docs is gpt-6-cyber.
OpenAI’s launch post includes named customer quotes. Firefox’s Brian Grinstead wrote that even after hardening with multiple frontier models, GPT-6 Cyber still found previously unknown high-severity issues and helped ship fixes. Calif’s Thai Duong called it “the best model we’ve used so far for long running exploit development tasks.” CodeRabbit’s Harjot Gill said it found 66% more high/medium security issues than their production model on an internal benchmark.
On Hacker News, the thread under OpenAI’s post was thinner on celebration and thicker on access design. One commenter asked the practical question many blue-team readers will ask:
“For defensive purposes, do I only need Daybreak Blue? I assume it’s much easier to get Daybreak Blue access than Daybreak Red.”
Another pressed on timing:
“What’s the point of delaying Daybreak Blue availability, when high-impact CVEs are being dropped several times a week?”
For: verified security teams already in (or applying to) Daybreak Red who need authorized exploit development and vuln validation at frontier capability — and who can meet the SSO, MFA, and IP-allowlist bar OpenAI documents for Red.
Not for: hobbyists, unpaid bug-bounty solo hunters without org approval, or anyone hoping this lands in ChatGPT Plus. Mainline GPT-6 Astra still covers ordinary code review and triage; Cyber is a restricted specialty model classified Critical for a reason.
Compared with open cyber-capable weights: labs like Moonshot and DeepSeek publish open models that refuse less on cyber prompts, but they are not the same product as a Daybreak-gated Critical model with OpenAI’s monitoring, PSP, and critical-infrastructure safeguards. Pick the stack that matches your authorization paperwork, not the leaderboard screenshot.
For coding-assistant context around OpenAI’s broader tooling, see our best AI for coding 2026 guide.
OpenAI: Introducing GPT‑6 Cyber ↗ — Nov 5, 2026
The Verge ↗ — Emma Roth, Nov 5, 2026