OpenAI released GPT-6 Cyber, its first model rated Critical in cybersecurity

On November 5, 2026, OpenAI introduced GPT-6 Cyber, a model trained to find and validate exploitable vulnerabilities while staying within the scope researchers set. It is initially available through the API and Codex to a select group of approved Daybreak Red customers, with access to expand to all Daybreak Red customers “soon.” OpenAI classifies it as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. At the same time, GPT-5.6 Cyber was opened to all Daybreak Red customers.

On ExploitGym, GPT-6 Cyber scores 48.5%, up from 32.1% for GPT-5.6 Cyber. On a new RemoteExploitBench, GPT-6 Astra under Daybreak Blue refused 100% of tasks while GPT-6 Cyber completed 95%. In the ExploitGym Honeypot test, which plants easy-to-exploit services outside the task’s scope, the model avoided them in 99.93% of cases without deployment safety prompts (GPT-5.6 Cyber: 94.49%) and 100% with them.

OpenAI says its researchers used the model to build 948 Linux privilege escalation exploits and identify 81 previously unknown high-severity Firefox vulnerabilities, many of which Mozilla has fixed. Named users include Cisco, Calif, CodeRabbit, Palo Alto Networks, CrowdStrike and SentinelOne. The API model page lists a 1,050,000-token context window and pricing of $20 per million input tokens and $100 per million output tokens, with Daybreak Red access and organizational approval required.

Why it matters: it is OpenAI’s first model it rates Critical for cyber, released a week after Google’s Gemini 4 Carbon took the same defenders-first route, which shows both labs now treating offensive cyber capability as something sold only to vetted customers.

What it does not show: ExploitGym, RemoteExploitBench and the honeypot test are OpenAI’s own evaluations, and the customer results are testimonials. The page does not say how many organizations have access or when broader access will arrive.