Hackers Are Stealing Claude Tokens From Subscribers

Hackers are stealing paid Claude usage from subscribers by hijacking login sessions and using compromised accounts to consume their AI token allowances, according to recent reports and warnings sent by Anthropic to affected users.

The issue came to light after Grant De Swardt, an independent AI consultant in the United Kingdom, noticed that his Claude Max 20x account was consuming tokens even though he was not actively using the service. After disabling connected tools and pausing scheduled tasks, his usage continued climbing. In one controlled period, his token consumption rose from 45% to 55% without corresponding activity from him.

After De Swardt contacted Anthropic, the company suspended his account, invalidated existing sessions and Claude Code tokens, and issued him a partial refund. Anthropic subsequently determined that a compromised Claude session key had been used to create unauthorized Claude Code OAuth tokens. The company said the evidence was consistent with either stolen credentials or session information being obtained without the user's knowledge, or the account having been connected to an outside service.

Anthropic has warned other affected users that a “bad actor” is using common infostealer malware to steal Claude login sessions from computers. The malware can capture saved passwords, session data and authentication information, allowing attackers to access accounts without necessarily needing the user's password again.

The reports suggest the problem may extend beyond a single account. After De Swardt shared his experience online, other Claude users described unexplained spikes in token consumption, including accounts that reportedly reached their usage limits despite little or no activity from their owners. Anthropic has identified and warned some affected subscribers, invalidated authorizations and issued refunds in certain cases.

One difficulty for users is visibility. Claude account support does not provide a detailed, itemized breakdown showing exactly what consumed a subscriber's tokens, making it harder to distinguish legitimate heavy usage from unauthorized activity. De Swardt said the lack of transparency made it difficult to determine how his account had been compromised.

The incident highlights a broader cybersecurity risk as AI services increasingly provide powerful coding and agent capabilities through paid accounts. A stolen session can give attackers access to valuable computing resources without requiring them to pay for their own subscriptions.

Anthropic did not provide TechCrunch with additional details on how users can identify misuse. For subscribers, the episode underscores the importance of protecting login sessions, avoiding suspicious downloads and monitoring unexpected changes in AI usage.