Stolen Claude Sessions Show the Risk in AI Subscriptions

An AI consultant in the U.K. noticed his Claude usage climbing on a day he was not working. After two weeks of back and forth, Anthropic told him a compromised session key had been used to create unauthorized tokens on his account. As reported by TechCrunch, he was not the only one — other subscribers described the same thing on Reddit and GitHub. If your company pays for AI tools, this is a spending and access problem you should look at now.

What actually happened

Grant De Swardt, an independent consultant who sets up AI agents for small and mid-size businesses, saw his token consumption rise on August 4 without doing any work. The next day he disconnected everything attached to his account and stayed off it. Usage rose again — in one controlled window it went from 45% to 55% with no active tasks running.

He asked Anthropic for an itemized breakdown of what was consuming his allowance. The company did not provide one, but agreed something was wrong. It suspended his paid account, invalidated his sessions and server-side Claude Code tokens, and refunded part of his $200-per-month subscription. The suspension knocked out his workflow, since he runs admin, coding and design work through AI agents.

Anthropic later told him a compromised session key had been used to mint unauthorized OAuth tokens, and that his account appeared to have been used by an outside service to handle activity for other people. It could not say how access was obtained. Two of the other affected users posted emails in which Anthropic said it had found a bad actor using common infostealer malware to lift Claude login sessions from people's computers, then using those sessions to burn through their usage. Infostealers sit on a machine and take saved passwords, session data and credentials. Anthropic said the malware did not come from using Claude — it comes from infected downloads, bad ads and similar sources.

Other users described usage jumping from 0% to 100% without touching the account, or from 0 to 49% in twelve minutes after a couple of prompts. De Swardt's account was reinstated after about two weeks. He cancelled anyway and moved to a tool that lets him switch between models. Anthropic declined to comment when asked how users can spot misuse.

Why it matters for a small business

Three things in this story apply to any company with paid AI seats.

First, a stolen session is not the same as a stolen password. Two-factor authentication does not help if the attacker copies a live session token from an infected laptop. The account simply keeps working, for them.

Second, you may have no way to see the theft. Support tracked total usage but not itemized usage, even on request. That means someone can quietly ride your subscription for a long time before anyone notices, and the only signal is a number that goes up faster than it should.

Third, the fix hurts. When the provider found suspicious activity, it suspended the account and invalidated sessions. That is the right security response, and it also means a business that runs its daily work through one AI vendor can be offline with little warning.

What to do about it

• Inventory your AI accounts and API keys this month. Write down who has access to each one, which machines are logged in, and which internal tools hold a key. Anything you cannot account for should be revoked.
• Stop sharing logins. One account passed around a team makes unusual usage impossible to attribute, and one infected laptop compromises everyone. Use separate seats or separate keys per person and per integration.
• Treat the endpoints as the weak point. Infostealers land through downloads and ads on ordinary work laptops. Keep machines patched, restrict what staff install, and be strict about browser extensions that can read session data.
• Set your own usage baseline. Note roughly what your team consumes in a normal week, and check the billing dashboard on a fixed day. A jump you cannot explain is the earliest warning you will get.
• Rotate keys and sign out sessions on a schedule, and immediately whenever someone leaves or a device is lost. Keep a written plan for what your team does if an AI account is suspended for two weeks.

The useful takeaway is not that one vendor is unsafe. It is that AI subscriptions have become a real line item and a real dependency, without the monitoring tools that come with older business software. Until providers show you what is consuming your tokens, your own records, tight access control and clean machines are the controls you have.