The Internet’s core does not need to know how it is used
AI agents are often described as a challenge requiring us to rebuild the Internet and to change how it is regulated. I disagree and in the following post I will explain why. The main point to stress here is that the shared core of the Internet, the DNS and the Internet Exchanges (IXes) where networks meet, cannot tell an agent from a browser. Indeed, these core functions do not need this distinction.
AI agents are likely to change where traffic goes, how many names are registered and how hard operators have to work on security. But these are questions of capacity and protection. The bigger risk is that AI ends up being regulated at an infrastructure level simply because that is where it is easiest to place a filter.
On 18 March 2020, a few days after much of Europe had been sent home to work due to COVID, my colleague Patrik Fältström wrote on this blog that “the increase in peak traffic is small, but the total amount of traffic per day is larger”. From where we sit at Netnod, that was more or less the whole story. A continent changed how it used the Internet within a week or so, and as far as I can tell nothing in the underlying infrastructure had to be redesigned. In fact very few people outside our field noticed, which is how it should be.
I used that example to open a short talk at EURid’s .eu Day in the European Parliament on 18 November, where the theme was the “autonomous web” and what AI agents mean for cybersecurity and digital trust. First, let’s clarify our definitions here: the “autonomous web” refers to a web where the majority of activity comes from software agents acting for people or organisations, with little or no human involvement at each step; AI agents describe software that acts on somebody’s behalf, again with little or no human activity in the loop.
My argument was that the shared core of the Internet should remain agnostic to AI and that we should be cautious about introducing AI-specific rules into the underlying shared infrastructure upon which everyone depends.
What the core sees
By the shared core I mean, roughly, the address book and the plumbing. The address book is the Domain Name System (DNS), which turns names such as netnod.se into the numbers computers use, from the root at the top down through the top-level domains (such as .se for Sweden). The plumbing is where networks connect to each other, such as at IXes. Netnod operates infrastructure across both areas including IXes in the Nordics, DNS services and i-root, which is one of the Internet’s thirteen root name server identities.
From the perspective of the Internet’s core infrastructure, a DNS query sent by an AI agent looks like a DNS query sent by a browser, and a packet sent by an AI agent crosses an IX like any other packet. There is no field in a DNS query or an IP packet that says “this is AI”, and there is no separate lane for AI agents (a web request can carry a label, but that is for the site at the other end to read). That follows from a choice made long before anyone talked about AI agents, which I discussed in an earlier post, “Nobody runs the Internet, and that was the point”.
By keeping the network simple and placing intelligence at the edges, the Internet has been able to accommodate successive waves of innovation from the web in the 1990s, streaming video in the 2000s and then smartphones without changing how its underlying infrastructure works. In this sense the core has never needed to know how it is being used, whether by a person sitting at a keyboard, watching a video or using a phone. An AI agent is simply the latest case.
Three ways AI agents could affect things
At EURid’s .eu Day at the European Parliament, the organisers asked me to look at three questions: what AI does to traffic at IXes and in the DNS; whether any infrastructure needs to change; and whether extra regulation is needed.
I do not believe AI agents require fundamental changes to the Internet’s underlying infrastructure. That does not, however, mean their impact will be negligible. I see at least three areas where they are likely to have an indirect effect.
The first is where traffic goes. For some time traffic has become increasingly concentrated among a small number of large operators. AI services are likely to reinforce this trend, with much of the processing taking place in large data centres and a growing share of traffic generated by machines communicating with other machines. At IXes this trend shows up as a few large networks carrying a growing share of the traffic. Cloudflare wrote in early November that agent traffic “may eclipse traditional human browsing on the Internet in the near future”. AI agents are also likely to look up many names that nobody has asked for recently, so that the answers are not cached nearby, which means more queries for the DNS.
The second effect will be on names. AI agents need addresses where they can be reached, and there are proposals (not yet adopted standards) for finding agents through the DNS. So we can expect to see more names, more frequent changes and an increasing reliance on automated systems to create and manage them.
The third effect will be on security, and this is the one I worry about most in my job. Attackers also use AI agents. The European Commission’s action plan on cybersecurity and AI from July describes how AI can be misused to find vulnerabilities and automate attacks at a speed and scale we have not seen before. Operators of core infrastructure are obviously among the targets.
What these three effects have in common is that they can be addressed through increased capacity and stronger security measures, rather than fundamental changes to the Internet’s infrastructure. These are challenges operators already deal with, although the additional security demands are likely to be expensive.
The dangers of regulating through infrastructure
The final question is whether AI requires additional regulation of the Internet’s underlying infrastructure. I do not believe it does. My concern is that attempts to regulate AI will increasingly target this infrastructure, simply because it offers a convenient point of intervention.
The temptation is easy to understand. If you want to stop a crawler, or decide which agents are allowed to do what, a filter is the obvious tool. But filters have a tendency to move downwards over time, from individual services to access providers and DNS resolvers (which look up domain names on behalf of users). Filtering at resolvers already happens, and a registry can act on a single name in its own zone. For example, Quad9, a public DNS resolver, announced that it would stop serving users in France from 24 November 2026, following a Paris court order requiring it to block certain domains in a sports-piracy case.
But what happens if the operators of the shared core, one step further down, are asked to enforce similar restrictions? These tools are blunt and, at the level of core infrastructure, could affect an entire top-level domain or network rather than a specific service or AI agent. Some operators, including root server operators, simply publish DNS data managed by others and have no authority to modify it. Nor can the underlying infrastructure reliably distinguish AI agents from other users.
This is not a limitation that needs fixing. It reflects a fundamental design principle of the Internet: its shared infrastructure was built to enable communication and innovation, not to enforce application-specific rules or determine which services and business models should succeed.
AI agents and the DNS: discovery, authentication and authorisation
When discussing AI agents and the DNS, it is important to distinguish between three functions: discovery, authentication and authorisation.
The first, finding an AI agent or the service it uses, is something the DNS can already support, just as it does for other Internet services. The second, verifying an agent’s identity, is a question of authentication. Work is underway in the Internet Engineering Task Force (IETF) on mechanisms for signing HTTP requests allowing services to verify information about the agents communicating with them. The DNS could play a supporting role here, for example by publishing cryptographic keys, much as it already does for email authentication.
The third, deciding whether an agent should be trusted or permitted to perform particular actions, is a matter for the service receiving the request, or for regulation directed at those operating the agent. It is this third function that I believe should remain outside the DNS. Turning the Internet’s address book into a mandatory register of AI agents would fundamentally change its role.
My recommendations to the European Parliament
At EURid’s .eu Day at the European Parliament, I ended with three requests to the policymakers in the room:
- Keep the shared core out of disputes about AI, and govern agents where they operate by regulating the conduct of those who run them.
- Judge security by whether services stay up under attack, and avoid pushing everyone towards the same solution since systems that are built the same way tend to fail at the same time.
- Remember the dependence that comes with all this, because an AI agent that runs in somebody else’s data centre is only there for you as long as your connection and your DNS work. So when public bodies buy connectivity or DNS for something that matters they should require redundancy in every layer. I should add that Netnod is one of the providers that sell such services.
Conclusion
The core of the Internet carries traffic from AI agents in the same way as everything else, and that simplicity is what has let it absorb every earlier shift, from the web to an entire continent working from home. Agents will probably concentrate traffic further, add names to the DNS and speed up attacks, and that calls for more capacity and more security work rather than new rules for the infrastructure. Decisions about whether to trust AI agents should rest with the services they interact with, while any regulation should target those who operate the agents rather than the Internet’s underlying infrastructure.
If your organisation is preparing to rely more heavily on AI agents, it is worth asking how well you understand the infrastructure on which these agents depend. As AI adoption grows, that question will only become more important. At Netnod, we will continue to monitor developments across our IXes and DNS infrastructure and share what we learn here.
Fredrik Lindeberg, PhD, Chief Information Security Officer, Netnod