GPT-6-Cyber launches with project approval and API IP restrictions

OpenAI added GPT-6-Cyber to its API on November 5, 2026, for investigating vulnerabilities, reproducing them in authorised test environments and checking fixes. Security teams evaluating it need to distinguish two boundaries: who may call the model, and which systems their agent is allowed to test.

The release is dated in OpenAI's API changelog, the same day as The Verge's originating coverage. The model documentation requires Daybreak Red access and approval for GPT-6-Cyber. Its appearance in the model catalogue does not establish access for every API organisation.

The API checks the caller and project

OpenAI's Daybreak setup guide says an organisation owner must enable the relevant access for the project making requests. With project-level controls, it specifies a non-default project and credentials belonging to that project.

GPT-6-Cyber API use also requires an allowlist of the public IP addresses from which the application sends requests. That controls where a request to OpenAI can originate. It is a different question from whether the application should be able to connect to a third-party server during an assessment.

For a team planning a trial, the immediate work is to confirm the exact approved organisation, project and model, then check the application's outgoing network identity. Approval for a related programme or another model should not be treated as a substitute.

The lab must enforce the target scope

OpenAI's recommended configuration says approved access and model safeguards do not replace an organisation's security controls. It calls for clearly defined systems, actions and time limits, backed by independent filesystem, network, identity and tool restrictions.

The practical implication is to design the evaluation around permitted work before attaching broad tools. A model that receives a written target list can still operate inside a poorly confined environment unless the surrounding controls enforce that list. Reachable services and connected tools belong in that assessment too.

Judge the workflow, not just a model label

TechKili's assessment is that a useful evaluation needs evidence of a reproducible finding, a valid fix check and respect for the approved scope. A successful API response alone establishes none of those outcomes. Nor do the documentation pages establish a measured rate of unauthorised actions in a particular deployment.

Our earlier coverage of Claude Opus 5.5's cybersecurity routing addresses another evaluation question: whether the workflow actually uses the model the team intended. Together, these checks help separate model access, execution behaviour and the quality of a security result.