EvilTokens: AI read 12,000 stolen inboxes and picked the fraud targets
A criminal service called EvilTokens charged $1,500 to join and $500 a month after that. For the money, it broke into company email accounts — more than 12,000 inboxes at more than 10,000 organizations — and then turned an AI chatbot loose on the mail to work out who inside each company was worth defrauding.
Microsoft took the service apart and published the details on September 22, 2026. Every fact and quotation below comes from Microsoft’s announcement, “Disrupting EvilTokens: The AI Chatbot Built for Cybercrime,” written by Steven Masada of Microsoft’s Digital Crimes Unit.
The story matters to you even if you have never thought about email security in your life, because of who got hit. Microsoft names the affected industries: wholesale distribution, construction, financial services, real estate, higher education, and healthcare. Not tech companies. Companies that send invoices and pay them.
What Microsoft actually shut down
EvilTokens launched in February 2026 and was sold through a Telegram channel for, in Microsoft’s words, “a $1,500 initiation fee and a recurring $500 subscription.” Within months it had been linked to “more than 12,000 compromised email inboxes across over 10,000 organizations worldwide.”
Microsoft says the heaviest victim activity was in the United States, followed by Canada, the United Kingdom, Australia, India, and France.
Working with partners and with authorization from the U.S. District Court for the Eastern District of Virginia, Microsoft seized 50 websites used to run the service and disabled more than 150 additional domains behind it. In the United Kingdom, the Metropolitan Police Service’s cybercrime team arrested two men, aged 32 and 38, on September 11, 2026, on suspicion of offenses connected with the alleged operation of EvilTokens. Microsoft says both men were released on police bail subject to conditions while the investigation continues. No one has been convicted of anything.
Microsoft calls this its Digital Crimes Unit’s 40th court-authorized disruption, and its first against an end-to-end AI-enabled cybercrime service. The platform is gone. The method it demonstrated is not, which is the whole reason Microsoft wrote the post.
The break-in was not a stolen password
This is the part worth slowing down for, because it is not what most people picture when they hear “hacked email.”
Nobody stole a password. Microsoft’s description: criminals got in “by tricking victims into entering an authentication code on Microsoft’s legitimate sign-in page.”
The code in question is a device code — the short string of characters a smart TV or a printer shows you when it needs you to sign in but has no keyboard. You type that code into the real sign-in page on your phone or laptop, and the TV gets signed in. It is a normal, legitimate feature.
The attack works like this, per Microsoft Threat Intelligence’s technical writeup:
- An employee gets an email with an urgent-sounding subject — Microsoft cites lures such as “Action Required: Password Expiration.”
- They click through to a page that generates a live device code on the spot and shows it to them, with a “Continue with Microsoft” button.
- That button sends them to the genuine Microsoft sign-in page. Real address, real padlock, real login box.
- They type in the code and approve. What they have actually approved is the criminal’s device, not their own.
- The criminal’s script has been polling in the background the whole time. The moment approval lands, the mailbox is open.
Two consequences fall out of that, and Microsoft states both plainly. First, no password was ever revealed, so nothing looked stolen — there is no “your password was changed” email, no obvious break-in to notice. Second, the access “could persist even after a password reset if the associated sessions and tokens were not also revoked.”
Read that second one again if you have ever handled a suspected compromise by telling someone to change their password. On its own, that does not close the door.
What the AI did once it was inside
Breaking into a mailbox is old news. What happens next has always been the slow part. Microsoft puts it this way: “Once inside an account, criminals have traditionally needed time and experience to sift through thousands of messages, identify decision-makers, understand payment processes, and find opportunities for fraud.”
EvilTokens handed that job to a chatbot. According to Microsoft, its AI tools could “summarize and translate emails, surface financial conversations, map organizational roles, identify trusted relationships, and recommend potential targets.”
It came with buttons for it. Microsoft lists the preset prompts the service shipped with — ready-made questions a paying customer could click rather than write:
- Find the wire-transfer discussions.
- Identify the organization’s “money movers.”
- Locate vendor invoices.
- Determine the best people to impersonate.
Microsoft’s summary of why that is different in kind, not just in speed: “In short, AI was not simply helping attackers write more convincing messages. It helped them decide who to target, who to impersonate, and how to most effectively exploit the relationship to extract as much money as possible.”
There is one more detail that says a lot about where this is heading. Microsoft’s investigators found that large portions of EvilTokens itself had been, in the report’s word, “vibe coded” — built with AI writing the software. The people running it did not have to be skilled programmers either. Microsoft also found the platform drew on capabilities from more than one AI model.
Why this reaches an ordinary business
Think about what is sitting in your controller’s inbox right now. Which vendors you pay. What your invoices look like. Who signs off on a payment over $25,000. How your bookkeeper talks to your best customer — the greeting, the abbreviations, the running jokes.
Everything a convincing fake needs is already in there. It used to take a person days of reading to extract it. Now it takes a question.
The payoff for the criminal is not a dramatic ransomware screen. It is one email, from a real address your team recognizes, in a thread they were already part of, saying the bank details for this month’s payment have changed. That email is correct about everything except the account number.
And the firms that got hit were wholesale distributors, construction companies, and real estate offices — businesses with a lot of money moving and not a lot of security staff.
What to do about it
None of this requires a security department. Most of it is a policy decision you can make this week.
- Verify every payment change by voice, on a number you already had. This is Microsoft’s own recommendation and it is the one that stops the fraud. Any request to change bank details, redirect funds, or approve an unusual transaction gets confirmed through what Microsoft calls “a trusted second channel” — a phone call to the number in your records, never the number in the email. Write it down as a rule so nobody has to be the one who decides to be suspicious.
- Teach one sentence: never type a sign-in code you did not ask for. If a page or an email hands your team a code and tells them where to enter it, that is the attack. Legitimate device codes only ever come from a device sitting in front of you that you are trying to sign in.
- When you suspect a compromised mailbox, revoke the sessions too. Resetting the password is not enough on its own — Microsoft says so directly. Whoever administers your Microsoft 365 needs to sign the account out of everything and revoke its tokens. Microsoft Threat Intelligence also recommends blocking device code sign-in entirely where your business does not need it.
- Assume minutes, not days. Microsoft’s advice is to “assume that once an inbox is compromised, criminals may understand its contents in minutes, not days.” A compromise you find on Thursday for something that happened Tuesday is already too late to treat as a small problem.
- Check your invoices for a while after any incident. Fraudulent payment redirects tend to show up weeks later, once a real invoice cycle comes around.
The takeaway
EvilTokens was a subscription product with pricing, customer support, and a dashboard. That is what the AI part really changed: the expertise that used to limit who could pull off this kind of fraud got packaged and rented out for $500 a month.
Microsoft has disrupted this one platform. The technique behind it is ordinary, and the next service to sell it will not be called EvilTokens. Your defense does not depend on catching them. It depends on a phone call before the money moves.
We break down AI news for business owners in plain language — more explainers in the AIWAS blog, and short alerts on X at @AIWASai.
Stop leaving customers on read.
We answer your reviews, keep your social pages posting, and make your ads convert, so you get more customers while you run the business.
Get Started