Who's Actually Running That AI Agent? A New Push to Give Them ID Cards

Vint Cerf, one of the two engineers who designed the protocols the internet itself runs on, left Google after twenty years this month to advise a company called Innovation Labs. The problem he's working on is a deceptively basic one: when an AI agent shows up on a website, an API, or another agent's inbox, there's currently no reliable way to know who it belongs to, or to hold that owner accountable if it does something wrong.

That sounds like a small gap, but it's the kind of gap that becomes a big problem the moment agents start acting on their own across the open internet instead of staying inside one company's app.

The specific problem: agents keep losing their identity

Right now, an AI agent's "identity" is usually just whatever credentials the platform it's running on happens to hand it. Move that same agent to a different platform, and it typically gets a new set of credentials and effectively becomes a different, unrecognizable entity — even though it's doing the same job for the same owner. Identity Digital, the DNS registry company behind this effort, points out that this makes it nearly impossible for anyone downstream to answer basic questions like "who deployed this thing" or "who's responsible if it makes a bad purchase or leaks data."

That's not a hypothetical concern. Identity Digital has cited research suggesting a large share of businesses now consider security, compliance, and auditability the main blockers to deploying agents more widely, and that a meaningful fraction of organizations that have already deployed agents have run into a security incident tied to one — frequently because basic access controls weren't in place.

The proposed fix: DNSid, a "birth certificate" for agents

Identity Digital's answer, built through its Innovation Labs division, is a standard called DNSid. The idea is to give each AI agent a persistent, verifiable identifier anchored to something that already has durable ownership records: a domain name. Combined with public-key cryptography, that identifier is meant to work like a birth certificate — proof of who created and owns a given agent that stays valid as the agent moves across different platforms and services, rather than resetting every time.

Crucially, Identity Digital has framed this as a neutral, non-proprietary standard rather than a product it controls: the company submitted the proposal as an Internet Draft to the IETF, the standards body that oversees core internet protocols, opening it up to public technical review rather than keeping it as a walled garden.

Why Cerf is involved, and why the comparison to TCP/IP matters

Cerf's own framing, in comments accompanying his advisory role, is that this is fundamentally a question of authority: what is an AI agent actually allowed to do, and how does anyone verify that after the fact. He's drawn a direct comparison to the early days of TCP/IP, the protocol he co-created — arguing that competing, incompatible approaches to a problem like this tend to get resolved not by one company winning, but by user and industry demand eventually forcing convergence on a shared standard.

That's a useful lens for a space that's currently crowded with competing efforts. Visa's Trusted Agent Protocol and Mastercard's Agent Pay, for instance, already verify agent identity — but specifically in the narrow context of a payment transaction, using their own card-network infrastructure. DNSid is pitched as something more general: an identity layer that isn't tied to any single use case, which in theory those payment protocols (or a customer-service agent, or a research agent) could all sit on top of instead of each building their own verification scheme from scratch.

What to actually watch here

A proposed IETF standard is a long way from something a random website checks before letting an agent through its front door, and plenty of DNS-anchored internet proposals never reach that point. The signal worth watching isn't the announcement itself but whether the large platforms actually building and hosting agents — the hyperscalers, the browser makers, the agent frameworks — pick it up, or whether they push a competing standard instead. Cerf's own bet is that it won't matter much which specific proposal wins, so long as some shared way to answer "who is this agent and who's accountable for it" ends up baked into how the agentic web works by default. Given how much of the current agent hype rests on trusting these things with real tasks and real money, that's a more foundational problem than it might sound like at first.