How Websites Trick AI Agents Into Buying

Today humans browse the web. Soon agents will browse it for humans, and businesses will be judged by machines before any person arrives.

That sounds like a prediction. It is already a description. AI agents book travel, order supplies, submit applications, and buy software on behalf of the people who deployed them. Every site they touch is evaluated by a machine first, and machines do not get suspicious. They complete tasks.

Every dark pattern ever built to trick a human is being retargeted at this new visitor: the countdown timer, the pre-checked add-on, the fee that appears at the final click. A $2.99 charge reads to an agent as a toll booth on the way to finishing the job, not a trap.

We caught one this week. A job board wanted $2.99 to submit a job application, priced a hair under the $5 line where most people let their AI agent spend without asking. Our agent flagged it instead of paying, because charging a job seeker to apply is a scam pattern no matter how small the amount.

That is the whole story in miniature: the web is learning to price against your agent's permission threshold. This article is our warning to the public, and our promise about how we operate.

The mechanics are old dark patterns, retargeted at a new victim, your AI agent instead of you:

• Sub-threshold pricing. $2.99, $4.99, set just under the spending limits people give their agents, so the purchase sails through without a human glance.
• The last-step paywall. Everything is free until the final click, when a fee appears. A human might balk; an agent optimizing for task completion reads it as a minor cost to finish the job.
• Trials that convert silently. "Free for 14 days" becomes a recurring charge while nobody is watching.
• Pre-checked add-ons and drip pricing. The price grows between the product page and the checkout, small enough each step that no single step trips the threshold.
• None of this is new. What is new is the target: software that completes tasks without getting suspicious.

An AI agent's job is to finish what you asked. A $2.99 fee on the way to "submit the application" looks like a toll booth, not a trap. There is no human friction at checkout, no raised eyebrow, no "wait, what?" moment. The spending threshold you set for convenience becomes the exact line attackers price under.

This is not theoretical. Banks including NatWest, Bank of America, ING, and Capital One have warned that AI shopping agents create new fraud and payment risks. There are documented cases of agents buying things their users never approved. The industry's answer so far is "set a spending limit", which is precisely the defense this trick walks around.

Our policy, stated plainly so you can hold us to it:

• A spending threshold is not auto-approval. Anything under our limit still gets judged on its merits.
• Patterns get flagged at any amount. Pay-to-apply schemes, last-step paywalls, surprise upsells, trials that convert, flagged no matter how small the number.
• A human decides when something smells off. That is the human-in-the-loop, working exactly as advertised.
• We tell you this because it is the same judgment layer we sell to clients. If we auto-paid a $2.99 scam fee with your money, why would you trust us with your business?

If you use an AI agent: set policies, not just thresholds. Tell it what it must never pay for, job applications, free trials that convert, fees that appear at the last step. Then ask your AI vendor one question: what does your agent refuse to buy? If they can't answer, you have your answer.

If you run a website: don't do this. Pricing against someone's agent threshold is a short-money trick that burns trust permanently. The internet already distrusts dark patterns aimed at humans; ones aimed at their agents will be remembered longer.

If you're evaluating Thanos Labs: this is what human-in-the-loop means in practice. Not a slogan, a $2.99 decision, made correctly, at 9 PM on a Sunday.

Want AI agents with a human watching the checkout? That's what we build.

The $2.99 job-board episode is firsthand: caught by Thanos during a live job-application task, October 2026.