Anthropic accounts compromised by infostealer malware to drain subscriber tokens
Article Summary
Anthropic has confirmed that a breach involving compromised Claude session keys led to unauthorized token usage and theft by a third-party service. Independent consultant Grant De Swardt noticed significant token consumption on his account without active use, prompting an investigation by Anthropic that resulted in his account suspension and a partial refund of £44.49. Similar incidents were reported by other users on platforms like Reddit and GitHub, with some accounts being auto-upgraded and rapidly depleted of tokens. Anthropic attributed these occurrences to infostealer malware stealing Claude login sessions from user computers, not from the Claude service itself, and has taken actions like signing out users and invalidating tokens for affected accounts.
AI Analysis
## Core Assessment
Anthropic’s recent issues with unauthorized token depletion stem from infostealer malware, which harvests active browser session cookies from user machines to hijack authenticated accounts. While Anthropic has correctly identified this as an external endpoint compromise rather than a service-side breach, their lack of transparent, itemized usage logging creates a critical visibility gap. For professional users—especially those who have integrated Claude into automated workflows—the inability to audit token consumption means they cannot distinguish between legitimate automated tasks and malicious siphoning. Even when the company intervenes by invalidating sessions, the process remains opaque, leaving victims to guess at the origin of the compromise while their operational continuity suffers.
## Impact and Advice
* **IT/Security Administrators:** Do not treat AI session tokens as disposable. Enforce endpoint protection to detect infostealer malware and strictly monitor for unauthorized browser extensions or suspicious OAuth authorizations.
* **AI Tooling Developers:** Implement granular, per-request usage logging. If a user cannot identify exactly what process triggered a token spike, they cannot effectively debug their own security posture.
* **Sole Proprietors/Power Users:** Audit your active API/OAuth connections regularly. If your subscription cost spikes without a corresponding increase in your own output, assume your session has been hijacked and force a global sign-out.
## Bottom Line
Anthropic’s platform currently provides no native way for users to verify what is actually consuming their quota, forcing victims to play detective against sophisticated malware. It’s a sobering reminder that as long as these services rely on long-lived session cookies, your “AI agent” is only as secure as the most vulnerable browser on your local machine.
Anthropic’s recent issues with unauthorized token depletion stem from infostealer malware, which harvests active browser session cookies from user machines to hijack authenticated accounts. While Anthropic has correctly identified this as an external endpoint compromise rather than a service-side breach, their lack of transparent, itemized usage logging creates a critical visibility gap. For professional users—especially those who have integrated Claude into automated workflows—the inability to audit token consumption means they cannot distinguish between legitimate automated tasks and malicious siphoning. Even when the company intervenes by invalidating sessions, the process remains opaque, leaving victims to guess at the origin of the compromise while their operational continuity suffers.
## Impact and Advice
* **IT/Security Administrators:** Do not treat AI session tokens as disposable. Enforce endpoint protection to detect infostealer malware and strictly monitor for unauthorized browser extensions or suspicious OAuth authorizations.
* **AI Tooling Developers:** Implement granular, per-request usage logging. If a user cannot identify exactly what process triggered a token spike, they cannot effectively debug their own security posture.
* **Sole Proprietors/Power Users:** Audit your active API/OAuth connections regularly. If your subscription cost spikes without a corresponding increase in your own output, assume your session has been hijacked and force a global sign-out.
## Bottom Line
Anthropic’s platform currently provides no native way for users to verify what is actually consuming their quota, forcing victims to play detective against sophisticated malware. It’s a sobering reminder that as long as these services rely on long-lived session cookies, your “AI agent” is only as secure as the most vulnerable browser on your local machine.