Redbot Labs / Threat Intelligence
EvilTokens Uses AI to Exploit 12,000+ Compromised Inboxes Across 10,000+ Organizations
- Incident ID
- RBT-TI-0087
- Organization
- Multiple Organizations Worldwide-EvilTokens Campaign
- Reported Location
- United States - Country-Level Victim Concentration / Global Campaign
- Severity
- High
- Category
- AI-Assisted Phishing / Business Email Compromise / Token Theft
- Record Date
- 2026-09-22
Incident Summary
Microsoft linked EvilTokens to more than 12,000 compromised inboxes across over 10,000 organizations worldwide following its emergence in February 2026.
The cybercrime service used AI to examine stolen mailbox access, identify trusted relationships and financial responsibilities, and recommend fraud opportunities. Its significance extends beyond generating convincing phishing messages: it helped attackers decide whom to impersonate and how to exploit existing business relationships. Microsoft disclosure
What Happened
Device-code authentication abuse
Phishing lures persuaded users to enter attacker-supplied codes on Microsoft’s legitimate sign-in page. Completing that process authorized an attacker-controlled session.
Token-based account access
Attackers used the resulting tokens to access mailboxes. Microsoft observed email extraction, malicious inbox rules, and—in some cases—new device registrations supporting continued access.
AI-assisted reconnaissance
The service analyzed compromised communications to identify valuable targets and financial discussions. Microsoft Graph access also supported organizational reconnaissance. Microsoft tracks the platform’s developer and support actor as Storm-2992. Microsoft technical analysis
Infrastructure disruption
Microsoft and partners seized 50 websites and disabled more than 150 supporting domains. UK police arrested two suspected operators; the investigation remained ongoing. Microsoft disruption report
Affected Systems
Organizational Email Accounts / Authentication Tokens / Inbox Rules / Device Registrations / Microsoft Graph Access
Attack Vector
Device-Code Phishing / Social Engineering / Token Theft / AI-Assisted Mailbox Analysis
Business Impact
Confidential communications exposed
Mailbox compromise can reveal commercial relationships, internal discussions, and information useful for follow-on attacks.
Financial impersonation risk
Access to genuine correspondence can make fraudulent payment requests more credible. AI-assisted analysis can help attackers identify the people and conversations most useful for impersonation.
Persistent access
Recovery may require investigation beyond a password change, including active sessions, device registrations, and mailbox rules.
Scope qualification
The reported organization count describes the reach of inbox compromises. It does not establish that every affected organization suffered a fraudulent payment or a broader network breach.
Redbot Analysis
Test the authorization decision behind the sign-in.
A familiar login page does not establish that the requested action is legitimate. Redbot recommends validating whether users can recognize which application or device they are authorizing and whether organizational policies restrict unnecessary authentication flows.
Assess what a compromised identity can reach.
Security testing should examine the account’s accessible resources, delegated permissions, and available actions. Use controlled test identities to determine whether an initial session can lead to sensitive communications or additional access.
Validate containment end to end.
An incident-response exercise should demonstrate that defenders can terminate unauthorized access, identify persistence, and confirm that the same access path no longer works. Record the time required to detect, investigate, and contain the test.
Protect payment approvals independently of email.
Business processes should require verification through an established channel before changing payment instructions. Test whether an attacker controlling a trusted mailbox could influence an approval without triggering that verification.
