Token Heist: Hackers Target High-Tier Claude Accounts to Steal API Usage

Independent tech consultant Grant De Swardt noticed something wrong with his Claude Max 20x account. His token usage kept climbing steadily on a morning when he had not opened the platform or run any local routines.

To isolate the issue, De Swardt disconnected every external service attached to his account. Despite shutting down scheduled tasks, local coding scripts, and cloud workflows, his token consumption rose from 48 percent to 68 percent during a short test window.

De Swardt contacted Anthropic support to ask for an itemized usage report. While customer service could not supply a detailed breakdown, support reps agreed that suspicious activity had occurred on the profile. Anthropic suspended his account, revoked existing session keys, and issued a partial refund for the remaining time on his $200 monthly plan.

Account suspensions disrupt small businesses that depend on automated software tools. De Swardt runs a small agency that helps local businesses set up coding agents to automate purchase orders and accounting data entry. As a solo founder, he relies heavily on smart assistants for web design, administrative work, and software development.

An internal investigation by Anthropic revealed that an attacker compromised a Claude session key to generate unauthorized Claude Code OAuth tokens. Support representatives informed De Swardt that someone used his account to process third-party requests, though engineers could not pinpoint how the attacker stole those credentials initially.

Because account dashboards report overall percentage usage rather than itemized activity logs, token theft can continue for months without subscribers noticing.

De Swardt shared his experience on Reddit, where dozens of subscribers reported similar issues. One user noted that their account auto-upgraded without permission, driving usage from zero to max capacity immediately. Another user recorded a usage jump from zero to 49 percent in under 12 minutes after entering standard web search queries.

A third user reported maxing out token limits for three consecutive days without touching the platform, leading them to open a detailed tracking report on GitHub where others shared matching stories.

Anthropic later sent warning emails to affected users, acknowledging that malicious actors used info-stealer malware to grab session credentials directly from personal computers. Info-stealers gather saved passwords, browser cookies, and local session tokens from compromised devices.

When Anthropic detected unauthorized access, security systems logged users out, revoked old authorizations, and issued partial refunds. Anthropic clarified that the malware did not originate from its core software platforms, noting that users often contract info-stealers by clicking infected online advertisements or installing compromised software packages.

De Swardt waited two weeks to recover his Claude profile. Frustrated by delayed customer support and a lack of itemized usage tracking, he canceled his Cursor subscription and moved his main workflow to open-source software options.

De Swardt noted that competing open-source models match proprietary software for everyday development tasks. He added that Anthropic still lacks native tools allowing users to monitor token consumption in real time, leaving subscribers vulnerable to credential theft.

When security systems lack granular usage tracking, subscribers struggle to protect personal credentials. Adding transparent activity logs helps software developers catch compromised session keys before attackers drain valuable account resources.