Security
GPT-5.6 Makes Frontier AI Access A Security-Control Problem
OpenAI's new account-security requirements for its most cyber-capable models show how frontier AI is moving from a product-access question to an identity and governance question.
By Leo W ·
OpenAI's decision to require stronger account security for access to its most cyber-capable GPT-5.6 models is a small product setting with a much larger implication: frontier AI access is becoming an identity-control problem.
The company says individual members who want to retain access to the most capable cyber models will need hardware-backed passkeys. That changes the frame. The question is no longer only whether a model should be released, or what it should refuse to do. It is also whether the person invoking it can be reliably tied to an account, a device and a security posture.
That distinction matters because powerful coding and security models lower the cost of legitimate defensive work while also reducing friction for abuse. A stolen password, a shared login or a disposable account becomes more consequential when the underlying tool can reason across code, infrastructure and vulnerability research.
Passkeys are not a complete answer. They do not determine whether a user has good intent, stop a determined insider, or remove the need for model-level safeguards. But they make it harder for an attacker to treat access to a sensitive capability as a commodity that can be borrowed, phished or quietly resold.
For enterprise security teams, the move is familiar. High-value systems already rely on phishing-resistant authentication, device trust, session controls and audit trails. AI vendors are beginning to import that discipline into a product category that was initially designed around fast, low-friction sign-up.
The harder work comes after login. Providers will need to decide what unusual activity looks like for an AI account: automated extraction, suspicious tool use, rapid changes in location, repeated attempts to skirt safeguards or behavior that indicates an account is being operated by someone other than its owner.
There is a tradeoff. More controls can be frustrating for researchers, developers and small teams that expect an API or chat product to work instantly. Yet the alternative is to pretend that a model with increasingly sophisticated cyber capability can be governed by the same account model used for an ordinary consumer app.
The industry has spent years debating model safety in abstract terms. GPT-5.6 makes the practical point: safety also lives in the account layer, where identity, recovery, logging and revocation decide who is actually holding the tool.
Topics: OpenAI, GPT-5.6, passkeys, AI security