Blog

What 40 Years of DNS Can Teach AI Agents

by Peter Lowe
What 40 Years of DNS Can Teach AI Agents
10 minute read

AI agents are growing faster than domain names ever did, without any of the registries, governance, or abuse handling that DNS grew up with.

Article in short: AI agents and DNS look alike on the growth charts. Both show geometric growth, a flood of standards and a lot of money. The difference is what sits underneath.

The Domain Name System (DNS) has registries, a governance body (ICANN), and formal ways to report abuse. AI agents have none of those yet, which makes problems like brand impersonation much harder to deal with.

Read on for what AI agents can borrow from DNS: a central registry, an ICANN-style governance body, and a real route for reporting abuse.

The ideas behind the Domain Name System (DNS) are 43 years old this year. Dr. Paul Mockapetris formalized them in November 1983 as two de-facto internet standards called Requests for Comments (RFCs): RFC 882, "Domain Names - Concepts and Facilities," and RFC 883, "Domain Names - Implementation and Specification." There are now by some estimates around 400 million domains registered, with 33 million net new domains added every year.

Pretty impressive for a system that's more than 40 years old and essentially the same as when it was first laid out — although of course there have been a lot of updates and addendums in the form of later RFCs.

Impressive, until you think about AI agents. ChatGPT was released in November 2022, so it's been around four years since the start of what most people think of as AI these days. Since then we've started seeing AI everywhere, with "agentic" AI the buzzword of the year in my subjective view. You could put that down to hype, but I don't think anyone disagrees that the AI and agents we're interacting with are here to stay.

We've been looking at AI agents for a while now, and as a self-professed DNS geek I think the parallels between DNS and the AI agent ecosystem make a comparison worth looking into. The number of proposals, registries, startups, standards and, well, AI agents has absolutely exploded, and there's a lot we can learn from how the DNS evolved and matured — because otherwise people are going to have to learn those lessons the hard way.

I'll go over some comparisons throughout this article. If you're mainly interested in what AI agents can learn from DNS, head to the lessons section toward the bottom. But the similarities (and differences!) are kind of fascinating.

How Fast Are AI Agents Growing Compared With Domain Names?

Zooknic has a nice list of the number of domains registered over time, going back to July 1, 1992, when there were roughly 16,000 domains. It took until 1997 to cross the 1 million mark. A lot of this is down to the popularity of the internet itself, but domain names are such a fundamental part of how it all works that it's impossible to separate the two.

AI agents, on the other hand, got up to speed much quicker. At ZeroFox, we monitor AI agents from all sorts of sources, and at the time of writing we're currently sitting at 735,000 agents discovered. It's hard to measure the growth exactly because there's no central system as with the DNS (more on that later), but one measure is GitHub repositories tagged topic:ai-agents. In 2023 there were 269. As of today in 2026, that's up to 124,000.

The takeaway: domain registrations grew at most 3.5x in a single year (1996 to 1997). GitHub repos tagged ai-agents grew 15.9x between 2025 and 2026.

Regardless of the exact numbers, you can see similar geometric growth — but with AI agents not only is the curve sharper, the timeline is much, much shorter.

There is of course a catch when measuring how many AI agents are out there, because there is no standard definition of "AI agent" itself, although most of us have a good idea of what it means. The bigger issue is that there's no centralized registry system like the DNS has. Registries will interact tightly with whatever standards we end up using, but what those will be is another matter.

How Many AI Agent Standards Are There?

When I first started looking into the number of proposed standards around AI agents, the number I came up with (OK, the number my LLM came up with) seemed... too many. I mean, just look at this graph:

The takeaway: DNS has produced 389 RFCs in 43 years. AI agents have 709 published standards and proposals after about three years, 658 of them IETF drafts with "agent" in the title.

How could there possibly be 709 standards and drafts related to AI agents already? But it's true — looking closer at the actual list, there really are that many. Here I thought the DNS had a lot! If you want to recheck this yourself, here's what the script did to compile the list from the Internet Engineering Task Force (IETF), the body behind most internet standards:

  • Get every draft with "agent" in the title from the IETF
  • Get each draft's first-version date: https://datatracker.ietf.org/api/v1/doc/newrevisiondocevent/?doc__name=<draft-name>&rev=00&format=json
  • Keep drafts from 2024 onward.
  • Remove non-AI uses of "agent" (user agent, relay agent, SSH agent).
  • Merge drafts with the same or near-identical titles.

Then we added proposed standards found through search and prior knowledge on top of that list. (Let me know if you'd like a copy of this output — I'd love to share.)

These are all drafts and proposals, and most of them won't make it much further. The list is also out of date the moment it's compiled, and there are probably some that have been missed — but it gives a good sense of the activity around standardizing AI agent behavior and the ecosystem in general.

How Much Venture Capital Is Going Into AI Agents?

Another metric that's revealing is the money. If you weren't around for the dotcom boom, it was shocking at the time to hear that someone had paid $10,000 for a domain, because online brand development wasn't really a thing for a while. Investment in internet companies skyrocketed — ultimately ending in a crash — but a lot of people made a lot of money, and the giants are still around today. The domain name and DNS industry is now a solid, reliable machine of its own that pays the salary of many people around the world.

This is absolutely dwarfed by the investment in AI agent companies:

The takeaway: DNS and domain startups peaked at about $266 million of funding in 2000 (in 2025 dollars). AI agent startups have raised $9.1 billion in 2026 so far. The shares differ in base: DNS peaked at about 0.2% of US venture capital, while AI agents are about 1.5% of global venture capital.

It's not even close. In the first draft of this post I was tempted to compare VC investment in any internet company, because the figures were more comparable — but that doesn't show the real picture. AI agents are only a subset of the whole AI industry (I'm not even trying to chart that out), just like the DNS and domain name industry is a subset of the dotcom boom, and the money is still astronomical.

What Can AI Agents Learn From DNS?

So what can we learn from all this? What made DNS successful was its elegant original design: hierarchical and scalable, able to adapt to exponential growth and resistant to massive horizontal scaling up. Right now, AI is moving so fast on iterative changes that the focus doesn't seem to be on long-term thinking, and that means a more unstable path ahead.

Why Do AI Agents Need Centralized Registries and Governance?

Currently, there are no centralized registries of AI agents, which means no internet-wide governance. Here's how centralized registration evolved in the DNS.

To be fair, the DNS was designed, so it was more naturally suited to evolving into the system we have today. Given the extremely rapid pace of developments in AI, though, there is arguably an urgent need for something like ICANN, the Internet Corporation for Assigned Names and Numbers, to come into play for AI agents. The multi-stakeholder model ICANN adopted can mean things take a while to change, but as a whole it has remained stable and inclusive, something most agree is a net benefit to the world.

The takeaway: DNS went from a single registry in 1983 to ICANN and a competitive market of registrars by 1998-99.

The lack of registries brings problems such as no visibility into who owns an agent, no standardized way to attach trust certificates, and nowhere to find transparency logs. These are all important, but the lack of governance has a consequence I’d consider more important —abuse.

How Do You Handle AI Agent Abuse?

This is what we've been researching in particular: abuse of, and via, AI agents. Online abuse is sadly inevitable, and agentic AI is no exception. At ZeroFox we're already seeing it in the form of brand impersonation, insecure exposed endpoints, untrustworthy Model Context Protocol (MCP) servers, and other threats you might expect.

AI agents are already empowered to talk to us and each other, recommend products, provide technical support, and make payments on our behalf. That makes them a perfect vector for abuse, and threat actors aren't foolish. Right now they're still working out how to monetize it at scale, but there is no doubt we're going to see more and more standardized attacks and brand abuse.

With a domain name and other kinds of DNS abuse, you can appeal to a registrar to get a domain taken down. There are established processes, channels and relationships to lean on, going all the way up to the Uniform Domain-Name Dispute-Resolution Policy (UDRP) for formal complaints. With an AI agent... good luck! I mean, OK, perhaps it's not quite that bad, but without a mature system it means abuse is much, much harder to deal with.

Compare this with the increase in phishing: even if we consider AI agents just another way to launch attacks, we're not ready for whats' about to happen.

The takeaway: the Anti-Phishing Working Group (APWG) saw about 50,000 phishing attacks in 2005 and 3.8 million in 2025, after a peak of 4.99 million in 2023 (APWG changed its counting method in mid-2020). Interisle counted 3.1 million phishing domains in the year ending April 2026, and 74% were registered by the attackers themselves.

The lack of governance means this is a hard problem to solve on a global scale, and it's a conversation ZeroFox is starting to have with multiple parties. If you want to get involved,let us know.

What the Future Holds for AI Agents

Not meaning to be hyperbolic here, but it's an incredibly exciting time. Things are fundamentally changing in real time around us, bringing improvements, problems, life-changing events and chaos in even measures. Predicting the future is always difficult, and anyone who tells you they know what's coming in the world of AI is almost certainly going to be wrong.

The DNS has shown us two things, though. First, a system designed and built with careful thought and consideration of the future can withstand the effects of geometric growth. Second, global coordination is necessary for world-changing technology. That's the lesson for AI agents and DNS alike.

Frequently asked questions

About AI Agents and DNS

DNS shows that a system designed with the future in mind can withstand geometric growth, and that world-changing technology needs global coordination. For AI agents, that points to centralized registries, a governance body in the mold of ICANN, and established ways to report abuse.

No. As of November 2026 there is no centralized registry of AI agents and no internet-wide governance. That means no standard way to see who owns an agent, attach trust certificates, or find transparency logs.

Nobody knows exactly, because there is no central registry and no standard definition of "AI agent." ZeroFox has discovered 735,000 agents from a range of sources. GitHub repositories tagged ai-agents grew from 269 in 2023 to about 124,000 in 2026, including 15.9x growth between 2025 and 2026. At its fastest, domain registration grew 3.5x in a year.

ZeroFox counted 709 AI agent standards and proposals in 2026 (to November 2), including 658 IETF drafts with "agent" in the title. Most are drafts and won't progress much further. For comparison, DNS has produced 389 RFCs in 43 years.

AI agent startups raised $9.1 billion in 2026 to October, up from $2.0 billion in 2023. DNS and domain startups peaked at about $266 million in 2000, in 2025 dollars. The 2026 figure is about 34 times the DNS peak.

For a domain, you can appeal to a registrar for a takedown, with formal complaints going up to the UDRP. AI agents have no equivalent, because there are no registries or governance to appeal to. ZeroFox is already seeing brand impersonation, insecure exposed endpoints and untrustworthy MCP servers.

ICANN, the Internet Corporation for Assigned Names and Numbers, was formed in 1998 to oversee DNS. It split the central registry from the registrars that compete to sell names. Its multistakeholder model can be slow, but it has stayed stable and inclusive, which is why a body like it is arguably needed for AI agents.

Sources: RFC 882, RFC 883, Zooknic domain counts, IETF Datatracker, ICANN UDRP. Chart sources are credited under each graphic: RFC Editor, GitHub search API, Tracxn, SEC and ICANN filings, US CPI, APWG, Interisle Consulting, National Science Foundation.

Peter Lowe

Peter is our Threat Research Lead in the Signals Research Program at ZeroFox, working on ways to improve our disruption efforts and coverage of the threat research landscape.

Peter’s career spans over 30 years, and he currently acts as the DNS Abuse Ambassador for the Forum of Incident Response and Security Teams (FIRST) as well as the co-chair of the DNS Abuse SIG. In his spare time, he runs one of the most popular blocklists for ads and trackers, and is active across a range of industry groups and efforts.