On 4 August Grant De Swardt, an independent AI consultant from East Sussex in Britain, noticed something strange on his paid account. He had not worked at all that day, yet his token usage was climbing by itself. The next day he shut down everything attached to the account and again did no work. The number went up again.

"In the cleanest controlled interval, it went from 45 to 55 percent while I did no work at all," De Swardt says. Tasks were paused or finished, cloud execution was off, not a single active local task was running. Somebody else was spending his subscription.

He asked for a list. There wasn't one.

Here the story gets more interesting than plain theft. De Swardt contacted Anthropic and asked for an itemised bill - what exactly was consuming his tokens. The company did not give him such a list, but agreed something was wrong. It suspended the paid account, revoked all sessions and all server tokens, and refunded him 44.49 pounds for the remaining time on a subscription of 200 dollars a month.

The suspension stopped his business. De Swardt makes a living setting up agents for small and medium firms - for instance, automatically moving order data into accounting software. He works through the same tools himself: admin, website design, coding. "Basically everything I do today goes through AI," he says. His account was returned after about two weeks.

When Anthropic finally identified the culprit, the explanation was that a compromised session key had been used to create unauthorised tokens. The company told him the account "appeared to be used by some suspicious third-party service processing activity for other people", but they could not establish how it got access. In other words: somebody got in and could quietly siphon off another person's plan for months, because support tracks total usage - but not what is using it.

That is the hole, and it is not a technical one.

De Swardt posted the case on Reddit and after 80 comments realised he was not alone. One user claims their account "was automatically upgraded without my consent, my card was charged, and usage jumped from 0 to 100 percent without me touching it at all". Another watched the percentage go from zero to 49 in twelve minutes after asking only a few questions and running one search. A third says his account burned through its entire daily quota three days running without him opening it - and filed a report on GitHub, where again people with the same problem turned up.

Two of them published emails in which Anthropic itself warned them. "We recently learned of an attacker using common information-stealing malware to take logged-in sessions from people's computers, then using them to access accounts and consume their quota," the message reads. The company logged users out, revoked authorisations, refunded part of the money and warned them they probably have an infection on their computer - which, it stressed, does not come from the platform itself but from everything downloaded and clicked online.

De Swardt received no such email. He says he found no evidence at all that his computer was compromised and that he still does not know how they got in. He cancelled the subscription and moved to a competitor's tool with cheaper models. "It's not that different or better," he says, adding that he cannot see himself returning until the problem is fixed. Asked how users are supposed to spot abuse, the company declined to comment.

What remains from all of this is one sentence worth more than the whole story: "I don't think these people have any way of protecting themselves." You pay 200 dollars a month for a service that measures your usage down to the percentage point but cannot - or will not - tell you what was using it. Every electricity, water or mobile internet provider in this region has to print you a line-by-line bill. Here that is still something you request as a favour rather than receive as a right.