Hackers Steal Tokens from Claude Subscribers via Infostealers

Grant De Swardt, an independent AI consultant from East Sussex, noticed something was off on August 4: his Claude Max 20x account was burning through tokens even though he hadn't done any work that day. The next day he disconnected everything that was hooked up to the service, paused or ended his scheduled tasks, and switched off cloud execution — and yet consumption still climbed from 45% to 55% without a single active task.

Unable to figure out what was eating his limit, he asked Anthropic for a detailed breakdown. The company didn't provide one, but agreed something was wrong: it suspended the paid account, revoked all Claude Code sessions and server tokens, and refunded part of his money — £44.49 for the remainder of the month on a $200 subscription.

A Compromised Session

After investigating, Anthropic reported that the cause was a stolen session key, which an attacker had used to issue unauthorized Claude Code OAuth tokens. According to De Swardt, the account “looked as though it was being used by a third-party service serving someone else's activity,” but exactly how the access was obtained could not be determined.

The crux of the problem is that support tracks only the total token spend, not a per-operation breakdown — even on request. That's why theft like this could go unnoticed for months.

He Wasn't the Only Victim

De Swardt described the case on Reddit and, in the 80 comments beneath it, discovered he wasn't alone. One user recounted that his plan was “automatically upgraded without consent,” money was charged to his card, and his usage jumped from 0% to 100%. Another's limit rose from 0 to 49% in 12 minutes, even though he'd only sent a couple of requests. Yet another reported on GitHub that his account had completely burned through the daily limit three days running without any input from him.

Two of them attached emails from Anthropic in which the company itself warned of the theft. “We became aware of a malicious actor using common infostealer malware to steal Claude login sessions from people's computers,” the email read. When suspicious activity was detected, Anthropic logged users out, revoked authorizations, partially refunded their money, and warned of a possible infection. The company stresses that the malware doesn't get in through Claude itself, but from third-party sources — infected software or ads.

De Swardt wasn't sent such an email. He claims he found no traces of compromise on his computer, and he has no way of determining how the access was obtained. His account was restored after about two weeks, but the incident and the lack of a spending breakdown cooled him on the service: he switched to Cursor, where you can use different models, including cheaper open ones. In his words, they work just as well.

When TechCrunch asked it to explain how users are supposed to detect abuse, Anthropic declined to respond.

Source: TechCrunch