OpenAI has announced that members of its Trusted Access for Cyber programme will need to use a physical security key to continue accessing its most capable cyber models from 1 September 2026.
Users will need to switch on OpenAI's Advanced Account Security feature using a hardware-backed passkey. The key can be tapped against a phone or plugged into a laptop. Yubico is supplying them, including a USB-C YubiKey and a low-profile key designed to remain plugged into a laptop.
Why identity teams should care
This is an interesting example of identity being used to control access to a capability rather than to data. OpenAI has decided the risk attached to these models is high enough that access should depend on a credential that cannot easily be phished or copied.
Most access control in the enterprise is still framed around records: which systems, which datasets, which customer accounts. Framing it around what someone can do is a different exercise, and it is the one that increasingly matters as tooling gets more powerful.
Expect this framing to spread. The question stops being "what can this person see" and becomes "what can this person cause".
The transferable question
Worth asking internally: which of your own tools would justify a hardware key requirement on capability grounds alone, regardless of the data they touch?
Likely candidates are production deployment pipelines, privileged access management consoles, identity provider administration, payment initiation, and anything that can issue credentials. If a compromised session on any of those would let someone cause irreversible harm, the credential protecting it should be as strong as the one protecting your most sensitive dataset, and often it is not.