What AI standards can steal from payments
Anthropic, Google and OpenAI want common AI safety tests, and payments already solved the hard part of writing them.
In brief
Anthropic, Google and OpenAI want common AI safety tests, and payments already solved the hard part of writing them. Agreeing a problem exists is easy. Standards only work when someone defines the exact test, the environment, and the threshold that counts as failure. Originally reported by Pymnts.
Agreement is the easy part
Anthropic, Google and OpenAI broadly accept that powerful models need common safety standards. The fight is over who writes the tests, who runs them, and what counts as failing.
Payments has been here
The card industry spent years turning vague trust questions into technical requirements. FIDO2 and WebAuthn landed in 2018 for passwordless login, EMVCo's common Contactless Kernel Specification in 2022.
Nobody certifies safe
EMVCo never declares card payments simply safe. It writes specs for defined pieces of the payment system, then sets up a process to check whether an implementation actually meets them.
“EMVCo doesn't certify that card payments are simply "safe."”
The tap that has to work
A terminal and a phone can come from different companies and still complete a transaction. EMVCo specs cover the exact technical requirements for that exchange, down to the kernel that processes it.
Define, specify, then test
The kernel spec names its functions: a secure channel, elliptic curve cryptography for card authentication, biometric and mobile verification, cloud operation. Then accredited labs test against it and approved kernels get published.
Vague categories are not standards
Demis Hassabis has proposed testing frontier models for dangerous capabilities, cyber risk and safeguard bypass. A real cyber eval needs defined tasks, tool and network access, an environment, and a failure threshold.
Who gets to act for you
FIDO's authentication standards replaced reusable passwords with public-key cryptography. Now its Agentic Authentication group is writing specs for how a user delegates actions to an AI agent with strong proof.
Registering an agent's intent
EMVCo's draft Agentic Payments Framework proposes an Intent Services layer where participants register and retrieve the instructions consumers give agents, plus possible know your agent checks. It is still only a consultation.
Specificity is the lesson
Payments bodies take a question like whether an agent was authorized and break it into data systems can exchange and verify. A cryptographic credential proves authority. Judging a dangerous model capability is far messier.
The disagreement travels with them
Anthropic backed Massachusetts AI safety rules while OpenAI opposed them and pushed Illinois-style annual third-party audits. Any new body inherits that split over evaluators, access, disclosure and revision speed.
“FIDO and EMVCo cannot tell AI developers where to draw those lines.”
What it means
Agreeing a problem exists is easy. Standards only work when someone defines the exact test, the environment, and the threshold that counts as failure.
