Can an AI agent be tricked into pressing Buy?

Yes, if its safety check only reads the button. We spent a day trying to get Nox, the agent on nox.cool, to press things without asking. Several tricks worked. This is what they were and what we changed.

Blog9 min readThe nox.cool team

A “Buy now” button with a pointer on it, stopped by a red no-entry sign, a green face beside it

Key points

  • An approval step is only as good as the check that decides when to ask. We tested ours by trying to get round it.
  • A check that reads the button's words misses a drag, a held key, a button in another language and a page that won't answer.
  • nox.cool now treats "I couldn't see what's there" as a reason to ask, never as "nothing is there".
  • After you say yes, Nox looks at the page again. If it changed while you were deciding, nothing is pressed.
  • Any button on a page that is taking a payment asks first, whatever the button says.
  • The words that ask (buy, pay, send, delete and more) are now known in the 37 languages the app speaks.
  • We replay a set of known attacks, 44 checks, before every release.

Can an AI agent be tricked into pressing Buy? Yes, if its safety check only reads the button. We spent a day trying to get Nox, the agent on nox.cool, to press things without asking you first. Several tricks worked. Here is what they were, why they worked, and what we changed.

Why does an AI agent need a check outside the AI?

Because the AI is the part that can be fooled. An agent that browses the web reads whatever a page puts in front of it, and some pages are written to mislead it. If the only thing between a hostile page and your money is the AI's own judgement, a clever enough page wins.

So on nox.cool the decision to ask you isn't made by the AI. Before Nox clicks, types or presses a key in its browser, the server looks at what that action would touch. If it's a button that pays, sends, deletes, posts or agrees to something, Nox stops and you get a card. The AI can't skip the card, because it never runs the check.

That design is sound. A design is one thing and the rule inside it is another, so we tested the rule.

How did we test it?

We read the rule the way an attacker would, asking one question of every line: what gets through? Then we tried each idea against a real browser and a local copy of nox.cool.

We weren't looking for ways to break into the server. We were looking for something quieter: an ordinary web page, or an ordinary sequence of steps, that ends with a button pressed and no card shown.

What got past the card?

Seven things did. None needed anything exotic.

1. A drag. The rule checked clicks. A drag that starts and ends on the same button also presses it, and a drag wasn't a click.

2. Enter, after some other keys. The rule checked what had the keyboard's focus before a key was pressed. Send "Tab, Tab, Enter" as one step, and the check looks at where focus was before the first Tab, not where Enter lands.

3. A held key. Holding Enter presses it many times. The first press was judged. The ones after it could land on whatever came up next, such as a dialog's Yes.

4. A page that won't answer. To judge a click, the server asks the page what's at that spot. A page can stall on purpose. When it did, the answer came back empty, and empty was read as "nothing there", which is harmless. The click went ahead.

5. A button inside a page component, or under an icon. Many sites build buttons out of their own components. Asked what was at the spot, the browser described the empty shell around the button, or the little icon inside it, neither of which says "Pay".

6. Another language. The words that made Nox ask were English. "Jetzt kaufen", "Acheter", "Pagar" and "购买" all mean buy, and none of them asked.

7. A change after you said yes. A card can wait for you for hours. When you approved, the press went to the same spot on the page without looking again. A page could show one thing while you decided and another when the press arrived.

What do those seven have in common?

Each one is the check judging something other than what actually happens.

It judged a click when the action was a drag. It judged the focus before the keys, when the press came after them. It judged the words on a shell when the button was inside. It judged the page as it was when you were asked, when the press came later.

That gave us one rule to rebuild around: judge the thing that will happen, at the moment it happens, and if you can't see it, ask.

What changed on nox.cool?

A press is a press, however it's made. A click, a drag, Space or Enter on a focused button are all judged the same way. Enter and Space must be sent alone, in their own step, one press at a time, so the check always looks at exactly what that one press will land on.

"I couldn't see" now means ask. If the page doesn't say what's at the spot within two seconds, Nox doesn't guess. You get a card that says it couldn't check what's there. An empty answer is no longer treated as an empty page.

Nox reads the real button. Its browser now looks inside page components and describes the button a click would press, even when the click lands on an icon or a word inside it.

The page is looked at again after your yes. When you approve, Nox checks that what's at the spot is still what you were shown. If it changed, nothing is pressed. Nox looks at the page afresh and, if the step is still right, asks you again about what's there now.

The words work in 37 languages. Buttons that mean buy, pay, order, send, publish, delete, confirm, subscribe and more now ask first in every language the app speaks. Words that mean something ordinary in another language were left out on purpose, so a page about Peru doesn't ask because "peru" means cancel in Finnish.

Context counts, not only the button. Some buttons say nothing: Continue, OK, Yes. So Nox now asks about any button on a page that is taking a payment, and any button in a dialog whose heading is about something that matters ("Delete this repository?"). A tick box asks too when its label opens something up, like Make public or Auto-renew.

Did we find anything besides clicks?

Yes. The same habit of mind turned up problems in quieter places. A few examples, all fixed:

  • A saved key could travel. You can save an API key for Nox's command-line tools and choose "Don't ask again" for it. That choice was meant for commands that reach only the key's own site. It didn't notice a site written without "https://" in front, and it didn't notice a command that turned the key back to front before printing it. Now it holds only for a plain call that hands the whole key to one program and reaches nowhere else.
  • Costs that weren't counted. When Nox hands a coding job to a coding tool, the job has a spending limit. The cost was counted when each answer finished, so hanging up a moment early counted nothing. It's now counted the moment an answer starts, and a job can have 6 requests going at once.
  • A sign-in code shown on the wrong screen. Signing in to the Mac app through your browser showed a code on both screens and asked whether they matched. Someone could have sent you their own link, and "yes, sign in" would have signed them in as you. Now the browser never shows the code. You type the one your own Mac shows, and 5 wrong tries end it.
  • Adding a way to sign in. Adding a wallet as a way into your account needed only an open window. It now needs a sign-in within the last 30 minutes, and you get a notice every time a way in is added or removed.

How can you see what Nox refused?

Open Receipts. It used to list what your agents did after you approved it. It now also lists what you said no to and what Nox refused to do, for the last 30 days, so a refusal is something you can see, not something that silently didn't happen.

The Receipts page with four views: All, Done, Said no and Refused, and a list of actions with what became of each
What was done, what you said no to, and what Nox refused, in the demo.

Will Nox ask more often now?

A little, and in the right places. You'll see more cards on checkout pages, in dialogs about deleting things, and on sites in other languages. You won't see more cards for ordinary clicks, links, typing or scrolling.

We take the trade the same way the rule always has: a card you didn't need costs one tap. A card you didn't get could cost money.

You can still tell Nox to stop asking about something specific, like one button on one site. A rule that covers a whole site now says in plain words that it includes buttons that pay, and for people who use Touch ID or Face ID for money it takes that check.

What does this say about AI agents in general?

Three things we'd pass on to anyone building or choosing one.

Ask where the check lives. If the answer is "the AI decides when to ask", the check can be talked out of asking. It should be a rule the AI can't reach.

Ask what the check looks at. Words on a button aren't enough. It should look at how the press is made, what the page around it is doing, and what's really under the pointer.

Ask what happens when the check can't see. The safe answer is that it asks you. The convenient answer is that it carries on.

None of this makes an agent impossible to fool. It makes fooling it harder, and it makes the failure a question on your screen where it used to be a press you never saw.

What do we do next?

We replay a set of known attacks, 44 checks, before every release, and it has to pass in full before anything ships. The tricks in this post have joined our tests. We'll run a review like this one again as nox.cool grows, because every new thing an agent can do is a new thing to check.

The other changes in this release, including answering Nox with a tap in Telegram and bringing back a deleted memory, are in the announcement. nox.cool is in early access, by invite for now. Open nox.cool.

Questions and answers

Can an AI agent buy something without asking me?

It depends on how the agent decides when to ask. On nox.cool, anything that spends money waits for your OK on a card, and the rule that decides when to show the card is enforced by the server, outside the AI. This post describes the gaps we found in that rule and closed on 11 November 2026.

What is an approval bypass in an AI agent?

It's a way to make the agent do something that should have asked first, without the question appearing. The usual causes are a check that only looks at one kind of action, such as a click, or one that only reads the words on a button.

How can a web page trick an AI agent?

By hiding instructions in what the agent reads, by changing what sits under the pointer after you approve, or by stalling so the agent's check can't see the button. The defence is to check the page itself, on the server, each time.

Does nox.cool ask before every click?

No. Ordinary clicks, links and typing go ahead. It asks before a press that pays, sends, deletes, posts or agrees to something, before any button on a page taking a payment, and whenever it can't tell what a press would do.

Do these checks work in other languages?

Yes. The words that ask first are known in the 37 languages the app speaks, so a button that says Jetzt kaufen, Acheter or Pagar asks the way Buy does.

How often does nox.cool test its own safety?

A set of known attacks, 44 checks, is replayed before every release, and it has to pass in full. The wider review described here was done on 11 November 2026.

Try it with Nox

Anyone can sign up: on the web, or with Nox for Mac.

Keep reading