🔴 When AI Becomes the Fraud Analyst

According to the supplied CyberRakshakLabs research, Microsoft described EvilTokens as an AI-enabled cybercrime platform combining account compromise, mailbox analysis, target selection and fraud preparation into a single service.

The article says Microsoft associated the activity with Storm-2992 and reported a Telegram-based subscription model with an initial $1,500 fee and $500 per month, with additional services reportedly available separately.

The important observation is the service model: expertise, infrastructure and workflow can be packaged so an attacker does not need to build every capability themselves.

The supplied research explicitly distinguishes the initial intrusion from the later AI-assisted activity. Microsoft reportedly said EvilTokens used device-code phishing to compromise Microsoft accounts and steal authentication tokens.

The supplied article stresses that the victim may never give the attacker a password; the authentication flow itself can be abused.

Once a mailbox is compromised, the attacker may face a large information problem. An executive inbox can contain invoices, bank details, payment approvals, vendors, contracts, executives and finance-team conversations.

The supplied research says EvilTokens could reportedly summarize and translate emails, identify financial conversations, map organizational roles and trusted relationships, identify payment authorizations, find vendor invoices, identify people involved in moving money, recommend impersonation targets and help draft fraudulent messages.

A compromised mailbox can contain the relationships that make business email compromise convincing: who reports to whom, who approves payments, which vendor is expecting money, when a payment is due and which employee should receive a request.

Traditional BEC can require compromise, manual reading, understanding the organization, finding a payment opportunity, impersonating an executive or vendor and redirecting money.

The supplied research describes the AI-assisted model as compromise → AI reads mailbox → AI maps organization → AI finds financial conversations → AI identifies target → AI helps prepare impersonation → fraud.

The critical change identified by the article is time: AI can potentially reduce the time between compromise and actionable intelligence.

That distinction matters because AI did not replace the initial intrusion. In the supplied analysis, AI compressed the post-compromise analysis and fraud-preparation workflow.

Traditional cybercrime required skill, time, infrastructure and knowledge. The supplied research frames AI-assisted cybercrime increasingly as:

The article places EvilTokens in the broader trend of Ransomware-as-a-Service, Phishing-as-a-Service, Malware-as-a-Service and Access-as-a-Service, describing an additional layer of AI-assisted Fraud-as-a-Service.

An executive mailbox can contain invoices, contracts, bank details, vendor relationships, payment approvals, employee and customer information, travel information, internal discussions, sensitive links and organizational structure.

The supplied research therefore frames the inbox as a business-intelligence database. Once compromised, the attacker potentially gains context—and context is what makes impersonation convincing.

The supplied research highlights independent verification as a business-process security control. Even if an attacker has the executive mailbox, vendor invoice, payment amount, employee name and perfect timing, Finance can interrupt the chain by independently verifying bank-account changes.

The supplied article says Microsoft reported a disruption involving a coalition including Health-ISAC, Cloudflare, Coinbase, OpenAI, Railway, SpyCloud, The Shadowserver Foundation and TRM Labs. It also reports the seizure of 50 websites and disabling of more than 150 additional domains, and says Microsoft reported two UK arrests on September 11, 2026.

Those claims are presented here as reported in the supplied CyberRakshakLabs article. The broader analytical point in that article is that disabling one platform does not remove the reusable model: compromise → AI analysis → target selection → impersonation → fraud.