What if your Claude account were quietly doing expensive work while your computer sat untouched? That is what one subscriber discovered after watching his token allowance rise despite not using the service. His experience has since uncovered reports from other users who say their Claude accounts were similarly drained without their knowledge.
• A Claude Max subscriber saw token usage increase while inactive
• The unexplained activity continued after connected tools were disabled
• Other users later reported similar unexplained token consumption
The incident began when an AI consultant in the U.K. noticed his Claude Max 20x account consuming tokens on a day when he had done no work. He repeated the test after disabling everything connected to Claude, pausing scheduled tasks and stopping cloud execution. His allowance still climbed, including one controlled period when usage rose from 45% to 55% without an active Claude Code task.
• The token usage continued under controlled conditions
• Connected services and scheduled tasks had been stopped
• The unexplained increase pointed to unauthorized activity
Anthropic eventually suspended the account, invalidated active sessions and Claude Code tokens, and issued a partial refund. Its investigation found that a compromised Claude session key had apparently been used to create unauthorized Claude Code OAuth tokens. The account appeared to have been accessed by a third-party service carrying out activity for other people, although Anthropic could not establish exactly how the attackers obtained access.
• A compromised session key was identified as the apparent entry point
• Unauthorized Claude Code tokens were created
• The precise method used to obtain the credentials remained unclear
The discovery became more concerning when the consultant shared his experience on Reddit. Other Claude users described accounts jumping from little or no usage to nearly full capacity without meaningful activity. Some users also shared communications from Anthropic warning that an infostealer may have stolen their login sessions. These types of malware can capture passwords, browser data and active session credentials, allowing attackers to access accounts without knowing the user’s password.
• Other users reported sudden and unexplained token consumption
• Anthropic warned some affected users about infostealer malware
• Stolen sessions can allow attackers to access accounts
The episode highlights a growing problem for AI subscribers: knowing that tokens were consumed is not the same as knowing what consumed them. Without detailed usage records, suspicious activity can remain hidden until a monthly allowance is exhausted. The consultant eventually regained access after about two weeks, yet cancelled his subscription. For Claude users, the bigger question remains whether there is enough visibility to spot token theft before the bill or allowance becomes the warning sign.
• Token theft may remain difficult to detect without detailed usage data
• Anthropic has invalidated sessions and issued refunds in some cases
• Users still need clearer tools to identify unauthorized AI activity
Via: Tech Crunch
