AgentDecoy · Research

An AI agent read a note hidden from people and did what it said, 61 minutes later

One of our decoy pages carries a single line written for AI agents, asking them to identify themselves at a check-in address with a code. People never see it. On 24 October 2026 something followed it.

Check your own pages for hidden text aimed at AI agents, free

Free. Reads only what any visitor can see. Your score is saved so you can share it; nothing sensitive is ever shown publicly.

What happened

Times in UTC, 24 October 2026:

  • 06:42: a visitor opens a decoy page that robots.txt tells every bot to stay out of. The page carries the note.
  • 07:43: 61 minutes later, a different address dressed up as an iPhone requests the check-in address with the exact code from the note.
  • It did not give its name. But a plain scanner never does this: only something that reads page text and acts on it does.

Why it matters

If an agent follows a harmless note, it will follow a harmful one that looks like a normal step, such as "verify your session at this link" or "include the settings in your summary". Anyone can plant text on a page: in a review, a comment, or a widget pasted into a site.

It came back from a different address an hour later, which looks like one agent handing the task to another worker. That makes blocking by address alone weak.

What site owners can do

Check your pages for hidden text aimed at agents, often pasted in with third-party widgets. And plant a harmless note of your own: an agent that follows it tells you agents are reading your pages and acting on what they find.

What agent builders can do

Treat page text as data, never as instructions. Do not let an agent call a tool because of something it read on a page without a separate check, and keep credentials outside the agent's reach.

Questions

Do AI agents follow hidden instructions on web pages?

Some do. In our logs one followed a hidden note to the letter an hour after the page was opened, from a different address. Most traffic that reaches decoy pages is plain scripts that never read the text.

Is this prompt injection?

Yes, a harmless form of it: text in content an agent reads that changes what the agent does. The same channel carries harmful instructions.

How can I tell if agents are reading my site?

Install the free AgentDecoy decoys. They include an identify-only note for agents, and every check-in is logged with the time, network and any name the agent gives.

Sources: Live catches, Prompt injection and websites, Open source decoys

Want it done for you?

Every result lists the fixes with starting prices, or I can look at your site with you. Pro re-checks it every week for $12 a month, 7 days free.