← Back to Tech · Lead story
Official OpenAI social graphic for GPT-6 Cyber on a dark background

OpenAI’s GPT-6 Cyber is a Critical-class model for authorized cyber defense

The defender kit just got a louder beep. OpenAI launched GPT-6 Cyber on Nov 5, 2026 — a cybersecurity-specialized model for authorized vulnerability research that the company now rates Critical on its Preparedness Framework, up from High for GPT-5.6 Cyber (OpenAI, The Verge).

What changed

  • Benchmarks: OpenAI says GPT-6 Cyber scores 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, and beats flagship GPT-6 Astra on most cyber tasks.
  • Staying in bounds: on ExploitGym Honeypot (built after the Hugging Face incident), GPT-6 Cyber avoided out-of-scope services in 99.93% of tests without the deployment prompt, and 100% with it — up from about 94.5% for the prior Cyber model.
  • Real systems: OpenAI reports the model helped surface 81 previously unknown high-severity Firefox vulnerabilities (Mozilla has shipped many fixes) and develop working privilege-escalation exploits for 948 Linux kernel findings in research campaigns.

Who gets it

Access starts with a select group of approved Daybreak Red customers via the API and Codex, with a wider Daybreak Red rollout “soon.” OpenAI is pairing the jump in capability with stronger misuse monitoring, IP allowlisting for Daybreak Red by default, and new checks aimed at blocking high-risk activity against critical infrastructure — while spinning up a Daybreak Red for Critical Infrastructure track for approved defenders.

Koffee take: a Critical-rated cyber agent is a big desk item for anyone who ships software. The headline that matters for normal humans is simpler — OpenAI is putting the sharpest knife behind a gated defender program, not a public free-for-all.

Filed Nov 5, 2026 ~8:20 PM CT — Evening pass. Sources: OpenAI — GPT-6 Cyber (Nov 5, 2026) · The Verge (Nov 5, 2026). Card image: official OpenAI GPT-6 Cyber social graphic.