RISKORIA ADVISING & PROFESSIONAL SERVICES D.O.O. · ZAGREB+385 97 737 1345INFO@RISKORIA.EU

Someone Else Is Spending Your AI Subscription, and You Have No Way to Know

10 SEPTEMBER 2026·11 MIN READ·TOMISLAV VAZDAR

Grant De Swardt did no work that day, and his Claude subscription was still spending someone else’s usage.

The UK-based independent AI consultant noticed on August 4th that his Claude Max 20x usage was climbing. He hadn’t triggered a single task. The next day he disconnected everything tied to his account and did no work at all. Usage rose again anyway. As TechCrunch reported, Anthropic later determined that a compromised session key had been used to mint unauthorized OAuth tokens for Claude Code. That let someone else spend his subscription without his knowledge. His account was suspended, every session invalidated, and he received a partial refund. It was reinstated after two weeks, but he cancelled anyway. Part of the reason was that Anthropic never handed him a breakdown of what had actually consumed his tokens.

De Swardt’s case turned out to be one data point in a wider campaign. Days earlier, The Register, Malwarebytes, Help Net Security, and BleepingComputer had each reported the same pattern. Usage refilled and then drained while account owners did nothing. Anthropic traced the campaign to well-known infostealer malware families. It named Vidar, LummaC2, StealC, RedLine, and Acreed on Windows, plus Atomic Stealer on a smaller number of Macs. None of these tools were built for Claude specifically. They are the same credential-stealing programs that have drained banking and gaming logins for years.

A stolen session is worth more than a stolen password

It is tempting to read this as one vendor’s problem. That is the wrong lens. This is what happens once a login stops being a simple password. It becomes a standing grant of compute, tools, and spending power.

The mechanism is worth spelling out. Infostealers do not crack passwords or defeat two-factor authentication directly. They copy the session cookie a browser already holds after login. That token proves both checks were already cleared. An attacker who replays that cookie is treated as the account owner, with no password prompt and no new-device alert. As far as the server is concerned, nobody logged in at all. That is why one victim burned through 49 percent of a Max plan’s monthly quota in twelve minutes with no visible warning sign.

A stolen banking password earns a fraud alert and a new card. A stolen AI session token hands an attacker a working agent that can execute code, call APIs, and run for hours unsupervised. All of it is billed quietly to someone else. What changed is not the malware, but the value now sitting behind that cookie. The attacker’s economics have changed. It used to take a skilled operator to do damage with stolen access. Increasingly, a stolen session and someone else’s agent are enough.

The gap in usage attribution

Anthropic deserves credit for part of its response. Once it detected the pattern, it signed affected users out and invalidated the stolen sessions. It removed saved payment methods so a hijacked account could not also be used to buy upgrades. It also refunded unauthorized charges. Its warning emails were direct about the source. They told users their machine was likely infected and that the malware had nothing to do with using Claude.

What stands out most in TechCrunch’s reporting is not the theft itself, but what happened to the users detection missed. De Swardt never received a warning email at all. He asked for an itemized breakdown of what had consumed his tokens, and the company could not provide one. Total usage is tracked. Per-task, per-session attribution apparently is not, at least not in a form support can hand back to a customer.

That gap is not unique to one vendor. Most organizations adopting agentic AI tools today have strong logging for their traditional identity stack. They have almost none for their AI usage stack. Every serious company can name who logged into its VPN at 3am. Very few can say which agent session spent last Tuesday’s API budget, and on what. Without that answer neither anomaly detection nor client accountability is realistic.

Treating AI credentials as privileged access

The practical response is not complicated. It does require treating AI session tokens with the same seriousness as admin credentials, rather than as a convenience layered on top of a subscription.

That means rotating and scoping tokens the way any privileged credential would be handled. It means not leaving long-lived sessions sitting in a browser profile indefinitely. It means running endpoint protection that actually catches infostealers. The malware has nothing to do with the AI tool itself and everything to do with hygiene on the machine holding the session. It means asking AI vendors now, before an incident forces the question, what usage attribution they can provide if something looks wrong. And it means monitoring consumption the way unusual login geography gets monitored. A sudden spike in token burn is the agentic-AI equivalent of a 3am login from a country where the company never operated. Riskoria builds this kind of monitoring into its incident management engagements. It treats agentic tools as part of the identity estate rather than something bolted on beside it.

What this means for your organization

The consultant in this story walked away from Claude mainly because nobody could give him clarity afterward. It was not because he was attacked in the first place. It is not really about which AI vendor is safest. Would your organization even notice if an agent’s credentials were being used by someone else right now? Could you explain what happened to a client or a regulator if they were?

Agentic AI is becoming core infrastructure inside businesses faster than the identity controls around it are maturing. That gap is exactly where the next wave of incidents will come from. It is also a gap Croatia’s Cybersecurity Act already covers. Essential and important entities must report unauthorized access to critical systems when it meets the criteria of a significant incident.

Source: TechCrunch.

ZKS COPILOT

A question about the Cyber Security Act? Copilot answers by measure and submeasure, with the article of the law.

Ask
SHARE THIS ARTICLE
NEW ARTICLES

One message every two weeks, with a link to the new article.

ABOUT THE AUTHOR

Tomislav Vazdar leads Riskoria, an advisory firm for cyber security, risk management, artificial intelligence and data protection. He has worked in the field for more than twenty five years and speaks at conferences and to the media.

Related reading

Want an opinion on your own situation, not a general text?

Half an hour, no obligation. Instead of a general text you get an answer for your organisation.

Arrange a callINFO@RISKORIA.EU · +385 97 737 1345