OpenAI puts GPT-5.5-Cyber into limited preview for vetted defenders

OpenAI announced on May 7, 2026 that GPT-5.5-Cyber is entering limited preview for defenders responsible for securing critical infrastructure and other specialized authorized environments.

The model sits inside OpenAI’s Trusted Access for Cyber program. OpenAI says normal GPT-5.5 remains the default for general work, GPT-5.5 with TAC lowers refusals for verified defensive workflows, and GPT-5.5-Cyber is the most permissive access level for specialized authorized work such as red teaming, penetration testing, and controlled validation.

OpenAI also says individual TAC users accessing the most cyber-capable and permissive models will need Advanced Account Security from June 1, 2026, while organizations can attest to phishing-resistant SSO.

Why this matters

This is not just another model picker entry. GPT-5.5-Cyber formalizes a split between general assistant access and gated, identity-verified access to more dual-use cybersecurity behavior.

For buyers, that creates a new procurement question: do you need a normal AI assistant, a verified defensive security assistant, or a highly controlled cyber model with stronger identity and monitoring requirements?

Buyer take

Most security teams should start with GPT-5.5 plus Trusted Access for Cyber, not assume they need GPT-5.5-Cyber. OpenAI says the Cyber preview is mainly about more permissive behavior for specialized workflows, not a blanket capability upgrade for every defensive task.

Critical-infrastructure teams should treat access as a governed security program. Require hardware-key or passkey readiness, scoped authorization, logging, and a clear policy for what systems the model may test.

What is still unclear

OpenAI has not published broad eligibility criteria, pricing, or general availability timing for GPT-5.5-Cyber. It also says future technical details from alpha testing will come through responsible disclosure, so the public still has only a partial picture of real-world performance.