Anthropic suspended a consultant's Claude account for two weeks to stop a token thief

The percentage on the usage screen is the entire artifact a Claude subscriber gets. It moves, and it does not say who moved it. De Swardt's cleanest measurement took his Max 20x meter from 45% to 55% on a day when scheduled Cowork tasks were paused or already finished and cloud execution was switched off ^[3]. That is ten points of a monthly allowance ^[1] spent by a party he could not name, on an account whose support tooling reports totals rather than line items ^[9].

Anthropic's account describes the mechanism in detail but does not explain how the credential was obtained. A compromised Claude session key was used to mint unauthorized Claude Code OAuth tokens ^[7], and to Anthropic the account looked as though an outside service had been running other people's activity through it, with no finding on how that service obtained access ^[8]. De Swardt says he found no evidence his own machine was compromised, and he never received the malware warning Anthropic sent to other affected users ^[15]. Those warnings describe common infostealer malware lifting Claude login sessions off people's computers and using them to consume the accounts' usage ^[13]. Anthropic signed those users out and invalidated their authorizations, issued some refunds, and said the malware had not come from Claude itself ^[14].

Fast theft is easy to notice. One Reddit user's account went from 0% to 49% in twelve minutes after a couple of prompts and a web search ^[11], roughly four points of allowance a minute ^[2]. Another burned its maximum three days running while the owner was not using it at all ^[12]. Anybody glancing at the meter would catch that. The case the missing log actually covers is the patient thief taking a slice a day inside a busy consultancy's normal variance, which per TechCrunch could continue for months before anyone thought to ask ^[9].

The containment lever and the outage are the same action. Anthropic suspended the paid account and invalidated every session and server-side Claude Code token ^[5], and the account came back about two weeks later ^[16]. De Swardt sets up agents for small and mid-size businesses, including ones that pull purchase-order data out of email into accounting software, and runs his own admin, website and coding work through agents as well ^[6]. The refund for the interrupted subscription was £44.49 ^[5], a defensible figure for unused time on a $200-a-month plan and unrelated to two weeks of a one-person business.

Then he left, and moved to Cursor for its multi-model support including cheaper open-source options, saying the alternatives were "not that much different or better" ^[17]. Substitution was cheap for him because he rated the capability as equivalent, so the thing that lost the account was the recovery path.

Anyone signing off on a metered agent tool faces two questions, both about the credential rather than the model. First, whether you can attribute spend to a specific session, key or machine from your own side of the account without opening a ticket. Second, whether you can revoke one compromised credential without suspending the whole account. If both answers are yes, a stolen cookie costs an afternoon. Claude subscribers currently sit with the first answer as no ^[9], which leaves a percentage as the only evidence in a dispute. When both answers are no, incident response looks identical to downtime; two weeks is what that measured out to here ^[16]. Teams that have already filled in those two boxes will spend an incident arguing about scope. The others will spend it asking support for a list that does not exist ^[4].