Verified, then logged.
Sun 22 Nov 2026 · 00:44 KST / 15:44 UTC한국어 | ENSearch /

01 / AI · Claim vs. Check · Analysis

OpenAI limits GPT-6 Cyber to approved security teams

OpenAI said the model helped find 81 Firefox vulnerabilities, but outside verification and a date for wider access have not been published.

An unbranded hardware security key is plugged into a laptop on a security team's desk, with an analyst seen from behind reviewing blurred code on monitors in the background. AI-generated image

OpenAI released GPT-6 Cyber on 5 Nov 2026 at $20 per million input tokens and $100 per million output tokens, for approved customers only, its model documentation says. Access runs through Daybreak Red, OpenAI's programme for cyber-specialised models, which applies identity verification and organisational security requirements.

OpenAI announced GPT-6 Cyber as a model for vulnerability discovery, exploit development and web penetration testing, and rated it Critical in cybersecurity and High in the biological and chemical domain. According to a Reuters report of the company's announcement, the model was used to find 81 Firefox vulnerabilities that Mozilla rated high severity.

OpenAI said it would widen access to all approved customers over several weeks. It has given no date.

Honeypot exploitation rate without deployment safety instructions (company-reported)

GPT-5.6 Sol49.5
GPT-5.6 Cyber5.51
GPT-6 Astra0
GPT-6 Cyber0.07
Source deploymentsafety.openai.com · as of 2026-11-05 · unit: Share of graded rollouts (%)Chart: made by LOGIT

A security variant of GPT-6 Astra

OpenAI released GPT-6 Astra on 3 Sep 2026, its changelog says. GPT-6 Cyber is a variant of that model with further training for security work, and it carries the same risk designations as GPT-6 Astra, according to the system card.

The system card said the further training used reinforcement learning on tasks such as reproducing crashes, developing working exploits and searching codebases for vulnerabilities. It also used reward signals designed to reduce refusals on authorised security requests.

The previous security model is GPT-5.6 Cyber. "In our evaluations, GPT-6 Cyber shows substantially stronger cybersecurity capabilities and improved alignment compared with GPT-5.6 Cyber," OpenAI wrote in the system card.

Strengths and weaknesses in the company's figures

The system card's company-published figures list 100% for general security assistance and vulnerability patching, 96% for exploit development tasks and 97% for red-teaming tasks, which test a system from an attacker's point of view.

Honeypot exploitation, without deployment safety instructions, is listed at 0.07%, against 5.51% for GPT-5.6 Cyber, and safe responses to severe-risk biology requests are listed at 99.87%, the same as GPT-6 Astra, the company said. The documents give no out-of-scope rate for real deployments.

Weaknesses in the company's own documents. In company-published figures, GPT-6 Cyber succeeded in 73.00% of tasks where stopping after a tool warning was appropriate, against 78.38% for GPT-6 Astra. In the blocker-reporting test, which grades the final answer, 20.29% of its responses were flagged, against 11.66% for GPT-6 Astra and 34.18% for GPT-5.6 Sol.

OpenAI said that in most flagged cases it reviewed, the model had already disclosed the blocker in an earlier message. "We evaluate these cases as incomplete final summaries rather than deliberative deception," the company wrote in the system card.

Who can use it and what it costs

Security teams. According to the system card, organisations must have single sign-on (SSO), multi-factor authentication (MFA), managed devices and least-privilege access. Individuals are limited to invited applicants who complete identity and liveness checks and hold at least one physical FIDO2 security key.

What is blocked. OpenAI said it added new safeguards to block high-risk activity against critical systems and assistance enabling such activity. It recommends that approved users run work in isolated environments, keep human oversight on high-risk workflows, and use sandboxing and Auto-Review when working through the Codex app.

OpenAI's documents list the Codex app and the API as access routes and require Daybreak Red access and approval for each organisation. They do not say whether organisations or individuals in Korea can obtain approval.

Cost. The table shows OpenAI's standard published prices, given in dollars only. GPT-6 Cyber prices are for short context with a context window of 1,050,000 tokens, and GPT-6.1 Sol prices are for prompts of up to 272K input tokens. GPT-6 Cyber cache writes cost $25 per million tokens for 30 minutes or $50 per million tokens for 12 hours.

ModelInput, per million tokensCached input, per million tokensOutput, per million tokens
GPT-6 Cyber$20$2$100
GPT-6.1 Sol$2$0.10$10
GPT-Rosalind$5$0.50$25

GPT-Rosalind is also limited to organisations approved through OpenAI's trusted-access programme, the changelog says.

Reproduction of the company-published figures by an outside body, an item-by-item list of the 81 Firefox vulnerabilities and Mozilla's own documentation have not been published. So far, every performance figure comes from the company.