OpenAI's cyber policy is not a locked door. It is a turnstile, and the turnstile is documented.

In April, OpenAI said it was scaling its Trusted Access for Cyber program to thousands of verified individual defenders and hundreds of teams responsible for defending critical software, and fine-tuning a cyber-permissive variant of its model, GPT-5.4-Cyber, for them. Access, it said, is shaped by clear, objective criteria, including strong identity verification, rather than by arbitrarily deciding who is worthy. [1]

That design annoys both easy arguments. X's open-access camp says defenders need the best tools immediately; the lockdown camp says cyber-capable models are inherently reckless. OpenAI's own record says the risk depends not only on the model but on who is using it, the trust signals around that user, and how much visibility OpenAI has into the activity. [1]

The developer documentation turns that principle into product behavior. It classifies GPT-5.3-Codex and newer models, including GPT-5.4 and GPT-5.5, as having High Cybersecurity Capability under the Preparedness Framework, which triggers automated safeguards in the API. If traffic crosses certain thresholds for suspicious activity, access can be temporarily limited and requests can return an error coded cyber_policy. [2]

The mechanics are specific, not rhetorical. For organizations without per-user safety identifiers, a flagged pattern can suspend the whole organization; for those that supply a unique identifier per end user, the suspension can be narrowed to the affected user after human review. For Zero Data Retention organizations, where OpenAI sees less, request-level mitigations apply on top, and the cyber_policy error can surface mid-stream. [2]

The Preparedness Framework supplies the logic underneath. OpenAI says it prioritizes capabilities that are plausible, measurable, severe, net new, and either instantaneous or irremediable, and lists cybersecurity among its tracked categories alongside biological and chemical capabilities. The claim is that a model reaching high capability ships only with safeguards meant to sufficiently minimize severe risk. [3]

The divergence is not whether cyber defense matters; everyone says it does. The fight is over who gets the strongest tools before the misuse case arrives, and OpenAI's answer is verification plus monitoring. Readers should judge that system by its receipts: who passes the turnstile, when access is revoked, how an appeal works, and whether a competent defender outside the favored circle can actually get through. The policy is published; the test is whether it is applied evenly.

-- THEO KAPLAN, San Francisco