GPT-5.5 and GPT-5.5-Cyber: OpenAI adds Trusted Access for Cyber

OpenAI expanded Trusted Access for Cyber with the release of GPT-5.5 and a specialized GPT-5.5-Cyber model, opening controlled model access to verified security teams and researchers. The move is designed to let vetted defenders perform vulnerability research and defensive testing with models and datasets that are normally restricted under standard safety rules.

The rollout pairs two model releases with an access framework. GPT-5.5 is a new baseline model in the GPT-5 line, while GPT-5.5-Cyber is presented as a domain-focused variant tuned or configured for cybersecurity tasks. Both are offered under Trusted Access for Cyber, a program that restricts use to approved organizations and applies additional oversight, contractual terms, and auditing.

What changed

OpenAI says the Trusted Access for Cyber expansion provides a way for verified entities to use advanced models for tasks such as vulnerability discovery, exploit analysis, and defensive automation. Access is gated by identity and eligibility checks, and operates under heightened monitoring and logging. The company emphasizes auditability and controls intended to reduce the risk that sensitive capabilities or data will be misused.

The cybersecurity variant, GPT-5.5-Cyber, is described as optimized for security contexts, including answering technical questions about network configurations, code-level vulnerability patterns, and exploit proof-of-concept reasoning. OpenAI pairs the model offering with usage controls that limit outputs and the contexts in which the model will operate. That includes restrictions on the types of payloads the model can generate and enforced human review for higher-risk queries.

OpenAI positions the program to serve a set of participants that includes corporate incident response teams, government defenders, and third-party security firms who pass verification and agree to the program’s terms. The company is also providing tooling for auditing, logging, and post hoc review of queries and outputs to ensure compliance with legal and operational requirements.

How access is structured

Trusted Access for Cyber is not an open endpoint. Organizations must apply, demonstrate a defensible need, and complete contractual and identity vetting. Approved users receive credentials for dedicated compute environments where models run within the program’s controls. Activity is recorded and subject to review by both the provider and, in some cases, external auditors or designated partners.

OpenAI frames the program as a tradeoff between enabling powerful defensive workflows and maintaining guardrails to prevent harm. The release notes and documentation emphasize limited sharing, detailed logging, and the possibility of human-in-the-loop checks for sensitive requests. Pricing and scale details for the cyber program are described as bespoke, with access tiers aligned to organization type and use case.

The announcement also outlines collaboration paths. Verified participants can request access to additional data sets or integrations that remain off-limits to general API users. OpenAI says these extensions are evaluated case by case and include contractual commitments around data handling and disclosure.

Why it matters

Opening controlled access to more capable models for verified defenders shifts some offensive-versus-defensive dynamics by lowering friction for legitimate security work while keeping stronger oversight than public endpoints. Organizations responsible for critical infrastructure and incident response teams could accelerate vulnerability discovery and remediation, but success depends on rigorous vetting, enforcement of usage limits, and transparent auditing. The program signals a move toward more fine-grained, use-case specific model deployments for high-risk domains.