Grant De Swardt, an independent AI consultant based in East Sussex, spotted a disturbing trend on August 4. His Claude Max account was burning through tokens despite zero active work. He paused all scheduled tasks and disabled cloud execution, yet usage climbed from 45% to 55%. When he demanded answers, Anthropic couldn’t provide an itemized log. They eventually suspended his account, invalidated his session keys, and refunded £44.49 of his $200 monthly subscription. But the damage was done. As a sole proprietor relying on AI agents for everything from coding to admin, the suspension halted his business cold.
Anthropic later identified the culprit: a compromised session key used to mint unauthorized OAuth tokens. The company blamed common infostealer malware, which steals saved credentials from infected computers. They warned other affected users that their login sessions had been hijacked by third-party services. Yet De Swardt insists his machine was clean. He found no evidence of malware. This disconnect highlights a terrifying gap in user control. Without granular usage logs, customers are flying blind. They can’t distinguish between their own prompts and a hacker’s silent siphoning.
De Swardt wasn’t alone. Reddit threads and GitHub reports flooded with similar stories. One user saw usage jump from 0% to 100% instantly. Another burned through daily limits for three days straight. Anthropic responded by signing out suspicious sessions and issuing refunds. But they didn’t send De Swardt the same malware warning email others received. He remained in the dark about the specific vector of attack. The lack of transparency frustrated him deeply. He couldn’t verify if his system was secure or if the breach was internal.
After two weeks, Anthropic reinstated his account. But the trust was broken. De Swardt cancelled his subscription, switching to Cursor for its multi-model support and open-source options. He argues that competitors offer comparable performance without the opacity. “I don’t think there’s any way that these people can protect themselves,” he said. Anthropic declined to comment on how users can better identify misuse. Until they build detailed audit logs, AI professionals remain vulnerable to invisible theft. The industry’s rapid adoption of agents has outpaced its security infrastructure.
