GPT-6 Cyber: OpenAI gates a Critical-rated cyber model

01What happened

OpenAI published the GPT-6 Cyber system card on November 5, 2026, on its Deployment Safety Hub. The card describes the model as a cyber-specialized variant of GPT-6 Astra designed for authorized security work, and states: "Like GPT-6 Astra, we classify GPT-6 Cyber as Critical in Cybersecurity under our Preparedness Framework." OpenAI says three outside parties took part in the evaluations, listed below.

02How it works

Staying inside the authorized scope is the design goal, and OpenAI measures it with a honeypot: a second, out-of-scope target planted in hard ExploitGym problems. Capability and restraint are scored separately, with intended-exploit rates on ExploitGym and a hardened variant on one side, and honeypot exploitation and task refusals on the other. Irregular ran external cyber evaluations using CyScenarioBench, FrontierCyber and Atomic Challenges, and Dragos co-developed a red-teaming evaluation for the industrial control safeguards in a controlled lab.

03Who is affected

Only approved organizations and individuals can call GPT-6 Cyber, because the card says access is restricted through Daybreak Red. Organizations must use SSO, MFA, managed devices and least-privilege access. Individuals must pass identity and liveness checks, show professional suitability and use at least one physical FIDO2 security key, and API IP allowlisting is on by default.

04What to do

Protect and rotate the credentials that can reach this model, which is what OpenAI requires of approved organizations, along with monitoring employee use and reporting suspected misuse. OpenAI says it also monitors for compromised API keys and access tokens on its side, and recommends Codex sandboxing, Auto-Review and human oversight for high-risk workflows.

Measured results

Questions people ask

Can I get access to GPT-6 Cyber?

Access to GPT-6 Cyber is restricted through Daybreak Red, OpenAI's program for cyber-specialized models. OpenAI says approved organizations must use SSO, MFA, managed devices and least-privilege access, while individual applicants must complete identity and liveness checks, show professional suitability and use at least one physical FIDO2 security key. API IP allowlisting is required by default.

Does GPT-6 Cyber attack targets outside the authorized scope?

GPT-6 Cyber exploited the planted honeypot in 0.07% of graded ExploitGym rollouts, by OpenAI's own measurement, against 49.5% for GPT-5.6 Sol and 5.51% for GPT-5.6 Cyber. GPT-6 Astra scored 0% on the same evaluation. Every one of these figures comes from OpenAI's internal testing, so read them as the lab's claim rather than an independent result.

How does GPT-6 Cyber compare with GPT-6 Astra at exploiting flaws?

The specialized variant is more accurate, not a new capability tier. GPT-6 Cyber reaches an intended-exploit success rate of 48.45% on ExploitGym against 42.43% for GPT-6 Astra, and 36.94% against 29.69% on the hardened variant. Both models carry the same Critical cybersecurity classification under OpenAI's Preparedness Framework, so the gate is the difference, not the rating.

Who checked OpenAI's cyber safety claims for GPT-6 Cyber?

OpenAI says it collaborated with NIST's Center for Advancing Innovation and Standards for Super Intelligence on measuring cyber capability, that Irregular ran external cyber evaluations using CyScenarioBench, FrontierCyber and Atomic Challenges, and that Dragos helped evaluate the industrial control safeguards in a controlled lab. OpenAI wrote and published the card, and the numbers in it are OpenAI's.

Sources 2 links, all fetched

1. 1deploymentsafety.openai.comGPT-6 Cyber Card
2. 2deploymentsafety.openai.comGPT-6 Astra Card