Skip to navigation

Threat reportPhishingTL-2026-2774

AI-Enabled Social Engineering and Synthetic Media (Deepfakes) Undermining Identity Verification

highACTIVE

Recorded Future's Insikt Group reports escalating AI-enabled social engineering: AI-enhanced phishing (malicious LLMs, EvilTokens device-code PhaaS, AI site builders) is largely countered by existing defenses, but synthetic media (voice cloning, live video deepfakes, biometric injection) erodes the audiovisual and biometric signals organizations use to verify identity. First disclosed .

Key facts

IdentifiersTL-2026-2774
Also known asSynthetic Media Social Engineering, Deepfake-Enabled Fraud
SeverityHIGH
First disclosed
Affected productsMicrosoft Microsoft 365 / Entra ID (OAuth 2.0 device authorization grant); Various Biometric identity-verification and liveness-check systems (KYC); Various Video-conferencing and voice-based approval workflows (e.g. Zoom)
Malwaretelegram, zgRAT
ToolsEvilTokens, FraudGPT, GhostGPT, KawaiiGPT, Tycoon, WormGPT4, Xanthorox
MITRE ATT&CK techniquesT1078.004 Valid Accounts: Cloud Accounts, T1087.004 Account Discovery: Cloud Account, T1098.005 Account Manipulation: Device Registration, T1114.002 Email Collection: Remote Email Collection, T1114.003 Email Collection: Email Forwarding Rule, T1204.001 User Execution: Malicious Link, T1528 Steal Application Access Token, T1539 Steal Web Session Cookie, T1550.001 Use Alternate Authentication Material: Application Access Token, T1564.008 Hide Artifacts: Email Hiding Rules, T1566.002 Phishing: Spearphishing Link, T1566.004 Phishing: Spearphishing Voice and 8 more in the attack flow
MotivationFINANCIAL
Target sectorsfinance, government administration, engineering, logistics, human-resources, cryptocurrency
Target regionsGlobal, North America, Europe, Asia
Detection rules9 (1 shown free)
Sources12 cited sources

Detection coverage

As of 2026-11-14, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-2774 across Splunk SPL, Microsoft KQL and Sigma, covering 40 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: Entra ID OAuth Device Code Sign-In (EvilTokens Device-Code Phishing) — Exploitation Phase

Detects successful Microsoft Entra ID sign-ins that used the OAuth 2.0 device authorization grant. In EvilTokens / Storm-2372 style campaigns the victim completes real MFA on Microsoft infrastructure after entering an attacker-generated device code, so the resulting device-code sign-in is the moment the attacker obtains OAuth tokens. Most tenants have very few legitimate device-code sign-ins, so any hit for a user who does not operate shared-input devices deserves triage.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1528, T1078.004, T1566.002, T1204.001
  • Data sources: Microsoft Entra ID SignInLogs (azure:monitor:aad), Microsoft Entra ID NonInteractiveUserSignInLogs (azure:monitor:aad), Sentinel SigninLogs, Sentinel AADNonInteractiveUserSignInLogs
  • Severity: high
  • Confidence: medium

Splunk SPL

``` Device-code sign-ins are rare in most tenants; success (errorCode=0) means a token was issued ```
index=azure sourcetype="azure:monitor:aad" (category="SignInLogs" OR category="NonInteractiveUserSignInLogs") properties.status.errorCode=0
| eval proto=lower(coalesce('properties.authenticationProtocol','properties.originalTransferMethod'))
| where like(proto,"%devicecode%")
| eval user=lower('properties.userPrincipalName')
| stats count as signins min(_time) as first_seen max(_time) as last_seen dc(properties.ipAddress) as ip_count values(properties.ipAddress) as src_ips values(properties.appDisplayName) as apps values(properties.location.countryOrRegion) as countries by user
| convert ctime(first_seen) ctime(last_seen)
| sort - signins

Microsoft KQL

union isfuzzy=true SigninLogs, AADNonInteractiveUserSignInLogs
| where TimeGenerated > ago(24h)
| where ResultType == 0
| where AuthenticationProtocol =~ 'deviceCode' or OriginalTransferMethod =~ 'deviceCodeFlow'
| summarize SignIns=count(), FirstSeen=min(TimeGenerated), LastSeen=max(TimeGenerated), IPs=make_set(IPAddress, 20), Apps=make_set(AppDisplayName, 10), Countries=make_set(tostring(LocationDetails.countryOrRegion), 10) by UserPrincipalName
| project UserPrincipalName, SignIns, FirstSeen, LastSeen, IPs, Apps, Countries
| order by SignIns desc

Sigma

title: Entra ID Successful Sign-In via OAuth Device Code Flow
id: 7c1e4a90-3b52-4f18-a6d3-0e9b2c815f47
status: experimental
description: Detects successful Entra ID sign-ins using the OAuth device authorization grant, the token-capture step of EvilTokens and Storm-2372 device-code phishing.
references:
  - https://www.microsoft.com/en-us/security/blog/2026/04/06/ai-enabled-device-code-phishing-campaign-april-2026/
  - https://abnormal.ai/blog/eviltokens-oauth-device-codes-bec-operations
  - https://www.recordedfuture.com/blog/ai-social-engineering
author: Threadlinqs Team
date: 2026-09-29
tags:
  - attack.initial-access
  - attack.credential-access
  - attack.t1078.004
  - attack.t1528
  - attack.t1566.002
  - attack.t1204.001
logsource:
  product: azure
  service: signinlogs
detection:
  selection:
    ResultType: 0
    AuthenticationProtocol: deviceCode
  condition: selection
falsepositives:
  - Legitimate sign-ins on input-constrained devices (Teams Rooms, smart TVs, CLI tools such as Azure CLI or az login --use-device-code)
  - Administrators and developers using device-code login for scripting
level: high

Known false positives:

  • Legitimate device-code logins from Teams Rooms, IoT/TV apps, Azure CLI, PowerShell or developer tooling
  • Admin scripting that uses az login --use-device-code

Tuning notes: Baseline device-code use for 30 days and allowlist known shared-device and CLI users. Better still, block the device code flow with Conditional Access where it is not required, and keep this rule as the tripwire. AuthenticationProtocol is missing from older tenants' logs; the query falls back to OriginalTransferMethod=deviceCodeFlow. Escalate when the source IP is a hosting/VPS range or the user has no prior device-code history.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • Endpoint Connection to EvilTokens, Device-Code Phishing, Lovable Lure and zgRAT Infrastructure — Exploitation Phase · unlock with TL Creds
  • Device Registration Following Device-Code Sign-In (PRT Persistence) — Post-Exploitation · unlock with TL Creds
  • Microsoft Graph Directory Enumeration and Payment-Thread Mailbox Mining — Discovery / Collection · unlock with TL Creds
  • DNS Lookups for Microsoft-Impersonating, Lovable-Hosted and Workers.dev Phishing Domains — Exploitation Phase · unlock with TL Creds
  • Device-Code Sign-In to Microsoft Authentication Broker / Device Registration Service — Post-Exploitation · unlock with TL Creds
  • Virtual Camera and Real-Time Face-Swap Tooling on Identity-Verification or Approver Hosts — Impersonation · unlock with TL Creds
  • Suspicious Exchange Inbox Rule Hiding or Redirecting Payment Mail — Post-Exploitation (BEC) · unlock with TL Creds
  • Entra ID Sign-In From EvilTokens / Device-Code Phishing Infrastructure — C2 / Token Replay · unlock with TL Creds

Indicators of compromise

As of 2026-11-14, Threadlinqs Intelligence tracks 40 indicators of compromise for TL-2026-2774.

Sample indicators (3 of 40)

  • IP address: 216.126.227[.]101
  • Domain: techroboticslabmade[.]com
  • URL: hxxps://app-54124296d32502[.]lovable[.]app/

Defanged, TLP:CLEAR. Full set of 40 indicators: unlock with TL Creds.

40 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

Documented losses include Arup (~$25M, Feb 2024) and a Singapore deepfake Zoom fraud (~SGD 4.9M, May 2026).

Insikt Group's 'Social Engineering in the Age of Synthetic Media' distinguishes two problems. First, AI-enhanced phishing: generative AI produces personalized lures, fraudulent login pages and automated follow-ups, and is packaged in underground offerings. The report names malicious LLMs (WormGPT, WormGPT4, EscapeGPT, FraudGPT, WolfGPT, DarkGPT, BlackhatGPT, KawaiiGPT, Nytheon, Xanthorox, GhostGPT, SheByte) and the EvilTokens phishing-as-a-service kit (emerged April 2026 per the report), which combines AI-generated lures, account validation, customizable infrastructure and automated creation of fresh device sign-in codes. The report also notes abuse of legitimate AI tooling, such as the Lovable AI website builder used to stand up large numbers of sites impersonating Microsoft, UPS and HR/financial platforms (August 2025), and open-weight models (GLM-5.2, Kimi, DeepSeek, Qwen, HUIHUI-AI reportedly used by the Gentlemen ransomware group). Insikt's assessment is that most of this can still be handled with existing defenses: email/web filtering, phishing-resistant MFA, monitoring for anomalous sign-ins, new inbox rules and unexpected permissions, and session revocation.

Second, synthetic media is the exception. Voice cloning of executives and officials, live video-call impersonation and falsified identity documents weaken the trust signals people and identity-verification vendors depend on. Facial biometric injection inserts AI-generated faces into video streams through virtual cameras or modified software, and screen-replay videos of synthetic blinking are used to pass liveness checks. Group-IB-recorded data cited by the report shows 8,065 biometric bypass attempts at an unnamed financial institution between January and August 2025; Telegram groups sell deepfake bypass tools aimed at platforms such as Binance, BBVA and Revolut. Human and automated detection is unreliable: a 2024 study put human identification accuracy of synthetic image/audio/video at 51.2%, and a May 2026 study found deepfake detectors only slightly better than chance.

Case studies: in February 2024 an Arup employee joined a video call with deepfaked CFO and colleagues and made 15 transfers to five Hong Kong accounts (~HK$200M / US$25M); no Arup systems were breached. In May 2026 a Singapore professional was contacted on WhatsApp by someone posing as the Secretary to the Cabinet, signed an NDA, joined a fabricated Zoom meeting with deepfaked senior officials, and transferred at least SGD 4.9M after a follow-up from a fake lawyer; Singapore Police cited poor lip-sync, distorted backgrounds, obscured Zoom logos and audio routed through a single account as tell-tale signs. The FBI has warned since 2023 (updated 19 Dec 2025) of smishing and AI voice-clone vishing impersonating senior US officials, moving victims to encrypted messaging apps.

Supporting technical reporting on EvilTokens (Microsoft, 6 Apr 2026; Abnormal, 3 Apr 2026; Sekoia, March 2026) shows the device-code flow being abused so victims complete real MFA on Microsoft infrastructure while the attacker captures OAuth tokens, then registers devices for Primary Refresh Tokens, creates inbox rules, performs Microsoft Graph reconnaissance and mines mailboxes for payment threads. The source report itself lists no IOCs and no nation-state attribution; the IOCs below come from the cited vendor reports.

MITRE ATT&CK techniques by tactic:

Initial Access

T1078.004 Valid Accounts: Cloud Accounts; T1566.002 Phishing: Spearphishing Link; T1566.004 Phishing: Spearphishing Voice

Discovery

T1087.004 Account Discovery: Cloud Account

Persistence

T1098.005 Account Manipulation: Device Registration

Collection

T1114.002 Email Collection: Remote Email Collection; T1114.003 Email Collection: Email Forwarding Rule

Execution

T1204.001 User Execution: Malicious Link

Credential Access

T1528 Steal Application Access Token; T1539 Steal Web Session Cookie

lateral-movement

T1550.001 Use Alternate Authentication Material: Application Access Token

Defense Evasion

T1564.008 Hide Artifacts: Email Hiding Rules; T1656 Impersonation; T1684.001 Impersonation

Resource Development

T1583.001 Acquire Infrastructure: Domains; T1583.006 Acquire Infrastructure: Web Services; T1584.001 Compromise Infrastructure: Domains; T1588.007 Obtain Capabilities: Artificial Intelligence

Reconnaissance

T1589.002 Gather Victim Identity Information: Email Addresses

Impact

T1657 Financial Theft

Remediation

Immediate actions

  • Require out-of-band verification (call a known number or use an approved system) for unusual, urgent or sensitive requests, including those apparently from senior leaders
  • Require an independent second approver for large payments, payroll changes, new access, account recovery and contact-detail changes
  • Revoke active sessions/refresh tokens and reset credentials after suspected device-code or AiTM phishing compromise
  • Hunt for unexpected inbox rules, new device registrations and unusual OAuth device-code sign-ins

Workarounds

  • Block OAuth device code flow via Conditional Access where it is not required
  • Apply Safe Links and anti-phishing policies; strong email and web filtering
  • Configure sign-in risk policies with conditional re-authentication

Longer-term hardening

  • Deploy phishing-resistant MFA (FIDO2 / passkeys)
  • Use deepfake detection only alongside independent verification procedures, not as a sole control
  • Deploy secure identity-capture with injection-attack and virtual-camera detection in KYC/liveness flows
  • Include current AI-generated lure examples in security awareness training and teach that polished, personalized messages are not proof of legitimacy

Affected and fixed versions

  • Microsoft — Microsoft 365 / Entra ID (OAuth 2.0 device authorization grant)
    Vulnerable versions: Tenants permitting device code flow
  • Various — Biometric identity-verification and liveness-check systems (KYC)
    Vulnerable versions: Systems lacking injection-attack detection
  • Various — Video-conferencing and voice-based approval workflows (e.g. Zoom)
    Vulnerable versions: Processes relying on audiovisual recognition for authorization

Timeline

Last updated · revalidated 1× · latest source

  • Arup employee transfers ~US$25M (HK$200M) in 15 transfers to five Hong Kong accounts after a video call with deepfaked CFO and colleagues (month-level date; publicly disclosed May 2024).
  • Start of the Jan-Aug 2025 window in which 8,065 biometric bypass attempts were recorded at an unnamed financial institution (Group-IB data cited by Insikt).
  • Proofpoint begins counting tens of thousands of Lovable-hosted URLs used for malicious activity (from February 2025), including a Tycoon PhaaS AiTM file-sharing campaign hitting 5,000+ organizations.
  • Storm-2372 device code phishing activity (February 2025), the precursor to the AI-enabled EvilTokens escalation per Microsoft.
  • Lovable-hosted UPS card-harvesting pages (ups-flow-harvester) first seen, exfiltrating data and SMS codes to a Telegram channel.
  • Lovable redirect to Dropbox-hosted RAR delivering DOILoader and zgRAT (C2 locked-ioc-20:7705) observed in a German invoice-themed campaign.
  • Proofpoint publishes research on cybercriminal abuse of the Lovable AI website builder for phishing, credential theft and malware delivery.
  • WormGPT4 malicious LLM identified (September 2025), sold on tiered pricing including a lifetime plan (day not stated).
  • FBI IC3 updates its PSA on smishing and AI voice-clone vishing impersonating senior US officials (activity since 2023).
  • EvilTokens PhaaS circulating in underground communities since mid-February 2026 (per Sekoia via secondary reporting).
  • Sekoia publishes follow-up analysis of EvilTokens; 1,000+ phishing domains identified by March 23 per CSA.
  • Abnormal AI publishes analysis of EvilTokens: device-code OAuth abuse, MailVault webmail client and AI-driven mailbox analysis.
  • Microsoft Defender Security Research publishes the AI-enabled device code phishing campaign report tied to EvilTokens.
  • Ahmedabad police disrupt a ring using replayed synthetic blinking videos to pass liveness checks and obtain fraudulent loans (April 2026).
  • Singapore Police advisory on deepfake Zoom scam impersonating the PM and senior officials; at least SGD 4.9M transferred by one victim (advisory dated 14-16 May 2026 across sources).
  • Open-weight GLM-5.2 release draws reporting on potential abuse by threat actors (Axios), cited by Insikt.
  • BleepingComputer reports the ARToken kit reuses EvilTokens device code API calls (POST /api/device/start).
  • LexisNexis reports a 180% year-on-year rise in deepfake attacks imitating live persons.

Update history

Sources cited

Previous and next in phishing

Drafted by the Threadlinqs research pipeline from 12 sources; published ; last updated .

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats