What if a cybercriminal could compromise an employee's Microsoft 365 account, analyze thousands of emails, identify your most valuable financial transactions, and create a convincing fraud attempt before your security team could respond?
That is the business risk demonstrated by EvilTokens, an AI-assisted phishing platform that Microsoft disrupted in September 2026.
The attack did not require a revolutionary vulnerability. Instead, criminals combined legitimate Microsoft authentication features, stolen access tokens, and AI-powered analysis to accelerate account compromise and financial fraud.
Key Takeaway: EvilTokens demonstrates how AI can accelerate familiar cyberattacks, reducing the time available for detection and response. Organizations need to prevent unauthorized access, restrict what compromised accounts and endpoints can do, and contain attacks before they cause significant business damage.
EvilTokens was a phishing-as-a-service platform that abused Microsoft's legitimate device-code authentication process to obtain access to Microsoft 365 accounts. Rather than simply stealing passwords, attackers tricked employees into completing authentication steps using legitimate Microsoft sign-in pages.
An employee could successfully authenticate using multifactor authentication (MFA) and still unknowingly authorize an attacker's session.
According to Microsoft Threat Intelligence's September 22, 2026 investigation, EvilTokens facilitated attacks compromising more than 12,000 inboxes across over 10,000 organizations.
After gaining access, attackers could use AI to examine mailbox contents, find valuable contacts and financial discussions, and develop business email compromise attempts. Microsoft identified the group behind the operation as Storm-2992.
AI allows cybercriminals to analyze stolen information, identify valuable targets, and prepare fraudulent communications faster and at greater scale.
Traditionally, attackers might spend hours reading mail to understand payment relationships and develop convincing impersonations. AI can accelerate that effort. Research from Huntress reported a 1,380% increase in observed device-code phishing attacks when comparing July through December 2025 with January through April 2026. Huntress examined EvilTokens activity across 344 victim organizations in a campaign wave and found no two phishing messages identical.
For businesses, the potential consequences include fraudulent payments, theft of confidential information, compromised accounts, and operational disruption.
Detection speed remains important, but attackers can accomplish more during the same response window when AI accelerates their work.
One attacker might gain access to a mailbox but accomplish little before being detected. Another might use automated analysis to identify payments and launch convincing fraud attempts before the same alert arrives. Both incidents could be detected in 30 minutes, but the outcomes could be dramatically different.
That is the concern in the October 26 CSO analysis. Security leaders must measure how effectively their defenses restrict attacker progress, not only how quickly they generate alerts.
Endpoint detection might catch suspicious device activity, but not every cloud account compromise requires malware on the victim's computer. EvilTokens abused legitimate authentication to acquire access to Microsoft 365.
An endpoint could appear healthy while an unauthorized session accessed cloud resources elsewhere. Identity controls, Conditional Access, phishing-resistant authentication, cloud monitoring, and session revocation therefore matter directly. Endpoint Detection and Response (EDR) remains important for identifying malicious processes and suspicious device behavior, but it cannot eliminate cloud account takeover risk by itself.
Yes. Security policies can restrict an attacker's necessary actions even when a specific threat has not yet been recognized. For EvilTokens, restricting unnecessary device-code authentication can prevent a method of unauthorized cloud access without recognizing the phishing campaign by name.
The same principle applies to Windows endpoints. Attackers may change malware, scripts, delivery methods, or AI tooling, but endpoint-based intrusions often still need to launch unauthorized processes, abuse legitimate applications, manipulate memory or files, establish persistence, or encrypt business data.
Changing the attack does not necessarily change the endpoint actions required to succeed.
Isolation and Containment technologies are designed to restrict those behaviors. AppGuard is an endpoint protection solution with more than a decade of production history focused on prevention through Isolation and Containment. It restricts unauthorized Windows endpoint behavior without requiring every threat to be identified first.
This can add protection against malware execution, ransomware, and certain post-compromise actions. AppGuard does not replace Microsoft 365 identity controls or, by itself, stop the cloud token abuse demonstrated by EvilTokens.
Review Microsoft 365 authentication policies. Evaluate whether device-code authentication is necessary and restrict or block it through Conditional Access where appropriate.
Strengthen identity and session protection. Use phishing-resistant authentication, review privileged accounts, restrict unmanaged devices, and monitor unusual token activity.
Prepare for compromised accounts. Establish clear procedures for revoking sessions, investigating suspicious inbox rules, and notifying affected contacts.
Add prevention beyond detection. Apply least privilege, application control, and endpoint Isolation and Containment where appropriate rather than assuming every attack will be detected in time.
Reduce the attacker's freedom to operate. Limit unnecessary access to sensitive data, administrative capabilities, applications, and critical systems.
Test complete attack scenarios. Simulate account takeover and determine whether your controls actually prevent financial fraud, persistence, and further access.
Verify payments independently. Require an established secondary channel for bank-account changes and unusual payment requests.
Yes. As AI reduces the time and effort needed to progress through an intrusion, preventive controls become increasingly important alongside detection and response.
EvilTokens did not invent phishing or authentication abuse. It showed how familiar techniques can become more scalable when AI accelerates them. The same lesson applies to malware and ransomware on Windows endpoints: changing the attack does not necessarily change the behaviors it must perform.
Unknown does not automatically mean unstoppable. The goal is not to predict every future attack but to restrict the actions attackers need to accomplish their objectives.
The critical question for business leaders is not simply how quickly their security teams can detect an attack. It is how much damage an attacker could accomplish before anyone realizes something is wrong.
Keep investing in Detect and Respond while strengthening identity security, reducing unnecessary access, and layering preventive controls that limit attacker behavior.
What if you did not have to detect every attack in order to stop the damaging actions it needs to perform?
Explore CHIPS cybersecurity resources and AppGuard endpoint protection for more on prevention-first security.