Grant De Swardt, an independent AI consultant based in East Sussex, U.K., encountered an unusual issue with his Claude Max 20x account on August 4. Despite not using the service that day, he noticed an unexpected surge in token consumption.
The following day, De Swardt took measures to disable all integrations associated with Claude, yet observed further increases in token utilization. “During a clearly controlled period, my usage jumped from 45% to 55% with no active tasks, scheduled tasks paused or completed, and all relevant services disabled,” he explained to TechCrunch.
Unsure of the cause, he reached out to Anthropic, the parent company behind Claude, for a detailed breakdown of his usage. Although they did not provide specific information, Anthropic acknowledged that something wasn’t right. They temporarily suspended his paid account, invalidated all active sessions and Claude Code tokens, and issued a partial refund of £44.49 from his $200 monthly subscription.
This suspension significantly disrupted De Swardt’s business operations. His work involves assisting small and midsize businesses in automating processes, such as transferring purchase-order data from emails to accounting software.
As a sole proprietor, De Swardt relies heavily on automated agents for various tasks, remarking, “Everything seems to be running through AI these days.”
Upon further investigation, Anthropic revealed that a compromised session key had led to unauthorized Claude Code OAuth tokens being created. The company indicated that De Swardt’s account had likely been accessed by an unknown third party, although the method of access remained unclear. “They suspect that either my credentials were stolen without my knowledge or that my account was linked to an outside service,” he recounted.
In essence, a hacker had infiltrated De Swardt’s account and was discreetly using his tokens. Due to the account’s tracking of total usage rather than detailed usage, this breach could have persisted undetected for an extended period.
After sharing his experience on Reddit, he discovered he was not alone. Others reported similar issues, with one user stating their account was auto-upgraded without consent, resulting in unexpected charges and usage spiking from 0% to 100%. Another user noted their usage surged from 0% to 49% within minutes, despite limited activity.
Some Claude users further detailed their experiences in a GitHub report, with two sharing communications from Anthropic that affirmed tokens had been compromised. “We have recently become aware of a bad actor using common infostealer malware to steal Claude login sessions from users’ computers, allowing unauthorized access to their accounts,” read the email. Infostealer malware can secretly gather saved passwords and credentials from affected devices.
Upon identifying suspicious activities, Anthropic took steps to log affected users out, invalidate existing authorizations, and issue refunds, while advising them about potential malware infections.
Importantly, the malware in question was not traced back to using Claude. Such threats can arise from various online interactions, including downloading infected software or clicking dubious ads.
Interestingly, De Swardt did not receive any alert from Anthropic, insisting there was no evidence of his computer being compromised and remaining uncertain about how hackers gained access.
After almost two weeks, De Swardt’s Claude account was reinstated. However, the experience of delay in resolution and lack of transparent usage tracking led him to terminate his subscription, opting for Cursor, which offers various models, including cost-effective open-source alternatives. He commented that other models perform comparably to Claude, stating, “It’s not that much different or better,” and expressed reluctance to return until Anthropic addresses the issues.
He emphasized that Anthropic still lacks tools to help users monitor token consumption, stating, “I don’t think there’s any way for users to protect themselves.” When approached for comment about how users could identify misuse, Anthropic chose not to respond.
Editor’s Take
This incident highlights significant vulnerabilities in AI subscription services, especially regarding user account security. As reliance on AI tools grows, so does the importance of robust and transparent security practices. For users and businesses, this case serves as a wake-up call about potential risks and the necessity for vigilant security measures to safeguard their digital resources.
Source: techcrunch.com
