For three decades, the internet has been built around one audience: people. Pages are designed for humans. For clicks ostensibly; so there's layouts, menus, and pop-ups that assume a person is reading them.
That assumption is now being tested as AI agents, software capable of browsing, reasoning, and completing tasks without anyone clicking a mouse, take over parts of that process.
Tools such as Perplexity's Comet browser and Anthropic's Claude browser plugin already attempt to act on a user's behalf, summarising pages, filling in forms, and booking services.
Reading pages differently
These agents do not read a page the way a person does; they parse its underlying code and any text on it, visible or not.
And this flip has already exposed a serious flaw in how the web is built.
In one documented test, a researcher hid an instruction in white text on a webpage, invisible to a human reader but fully visible to an AI agent.
When asked simply to summarise the page, the agent instead followed the hidden instruction and began composing an email to an address it had never been given by the user.
In separate tests, browsing agents have acted on instructions buried inside emails, including opt-out requests and meeting details, without flagging them to the person who asked for a summary.
Two versions of the same page
The agent cannot easily tell the difference between an instruction from its user and one planted by whoever built the page or wrote the email.
This is possible because agents and humans are, in effect, reading two different versions of the same internet.
A person sees a rendered page: images, formatted text, buttons in familiar places.
An agent sees raw code, structured data, and any text embedded in that code, whether or not it was ever meant to be seen.
A widening gap
As more software reads and acts on that hidden layer, some commentators argue the internet is developing two audiences instead of one: humans clicking through visible pages, and machines parsing structure and instructions underneath.
Automated, non-human traffic already accounts for a growing share of activity online, and some analysts now describe a shift toward machine-generated content outweighing human-authored pages entirely.
Practical issues
These are not abstract problems. If agents cannot reliably distinguish a genuine user request from an instruction hidden in a webpage or email, then the parts of the internet built without humans in mind become an attack surface. In other words, an efficiency hack becomes a security problem. A massive problem.
