Hackers Are Stealing Claude Tokens From Anthropic Subscribers

Anthropic Claude subscribers are reporting unauthorised token usage after attackers accessed accounts through compromised sessions and used them without the owners’ knowledge.

The issue came to light when AI consultant Grant De Swardt noticed his Claude Max 20x account consuming tokens even when he wasn’t actively using the service. After turning off connected tools and pausing scheduled tasks, his token usage kept increasing.

De Swardt contacted Anthropic for an explanation. The company suspended his account, invalidated sessions and Claude Code tokens, and provided a partial refund for his subscription after determining that unauthorised activity had taken place.

Compromised Sessions Used to Access Claude Accounts

Anthropic told De Swardt that a compromised Claude session key was used to create unauthorised Claude Code OAuth tokens. The company said an unauthorised third-party service appeared to have accessed the account, although it could not determine exactly how it obtained access.

The incident suggests attackers may be targeting AI service accounts in the same way they target other online accounts, using stolen sessions or credentials rather than directly attacking the underlying AI systems.

Because Anthropic does not currently provide detailed, itemised token usage reports to subscribers, users may struggle to identify unauthorised consumption until a large amount of usage has already occurred.

Other Claude Users Report Similar Problems

After sharing his experience online, De Swardt found other Claude users reporting similar cases of unexpected token consumption.

Some Claude users reported on Reddit that their Max subscriptions reached full usage without normal activity. Other users described sudden increases in token consumption despite limited or no interaction with Claude.

A separate GitHub discussion about Claude Code also included reports from users investigating unexpected account activity.

Anthropic Warns About Infostealer Malware

Anthropic has warned some affected users that attackers may have used infostealer malware to capture Claude login sessions. This type of malware can steal saved passwords, browser sessions, and other authentication data from infected devices.

The company said the malware was not caused by using Claude itself and could originate from other online activities, including downloading infected software or interacting with malicious advertisements.

When it detects suspicious activity, Anthropic said it can sign users out, invalidate authorisations, and provide refunds in some cases.

Security Concerns Grow Around AI Accounts

The incidents highlight new security challenges as AI services become more integrated into professional workflows. Many users now rely on AI tools for coding, business operations, research and daily administrative tasks.

For users running AI agents or automated workflows, unauthorised access can consume expensive subscription limits and disrupt important work.

Anthropic has not publicly explained every detail of the reported incidents, including how attackers obtained some compromised sessions. The company continues to advise users to protect account credentials and monitor unusual activity.