The AI gold rush has created a new social engineering surface, and threat actors are exploiting it at scale. Between April and October 2026, ANY.RUN linked 1,258 public sandbox analyses to campaigns impersonating popular AI products — fake installers, cloned landing pages, and malvertising lures that push malicious software under the branding of tools employees are actively asking their IT departments to approve. US businesses accounted for 16% of country-attributed targets, making this a domestic problem, not an overseas curiosity.
The payloads observed in these campaigns follow a consistent pattern: infostealers and loaders that steal browser sessions, harvest stored credentials, and pivot into business account compromise. For MSSPs and in-house SOC teams, this is a high-volume, low-sophistication threat with high-impact outcomes — a single stolen session cookie can bypass MFA entirely and give an attacker a authenticated foothold in M365, Google Workspace, or SaaS platforms within minutes of execution.
This post breaks down the attack chain, provides deployable detection content (Sigma, KQL, VQL), and lays out the containment and hardening steps your clients need in place before the next wave of lures lands.
Technical Analysis
The Attack Chain
These campaigns are not exploiting a software vulnerability — they are exploiting user trust in AI branding and the gap between employee demand for AI tools and IT approval workflows. The typical chain:
- Delivery: Malvertising (poisoned search ads for terms like "AI assistant download"), SEO-poisoned cloned sites, sponsored social posts, and lookalike domains (typosquats and homoglyph variants of legitimate AI vendor domains) push victims to download an "installer."
- Staging: The installer is typically an unsigned or fraudulently signed executable (often an MSI, or an EXE wrapped in a legitimate-looking installer framework like Inno Setup or NSIS) dropped to
%USERPROFILE%\Downloadsor%TEMP%. Many samples observed in sandbox analyses stage additional payloads from attacker-controlled CDNs after an initial "decoy" installer window displays, preserving the illusion of a working product. - Execution & theft: The payload is most commonly an infostealer-family binary. It reads browser credential stores and session data — Chrome/Edge
Local State,Cookies,Login Data, and theLocal Extension Settingsdirectories — decrypts them using DPAPI via the browser's own encryption keys, and exfiltrates over HTTPS. - Monetization: Stolen session cookies are replayed to hijack authenticated sessions (MFA bypass via token theft), and harvested credentials are sold or used for direct business account takeover, BEC, and follow-on intrusion.
Why This Works Against Businesses
- AI tool sprawl: Employees install AI tools faster than procurement can vet them. Shadow AI adoption is the delivery mechanism.
- MFA bypass via session theft: Cookie replay defeats most phishing-resistant postures that stop at password+OTP. Only token binding (e.g., Microsoft token protection/Conditional Access session controls) reliably blunts it.
- Low detection coverage: Stealer execution happens in user context, often as a child of a user-launched installer, and completes in seconds. Signature-only AV coverage is inconsistent against repacked samples.
Exploitation Status
No CVE is associated with this activity — it is pure social engineering plus commodity malware tradecraft. The campaigns are confirmed active in the wild across a six-month window (April–October 2026 per ANY.RUN's corpus of 1,258 public analyses), and the volume trend indicates continued operation. Treat any AI-branded installer outside your approved software catalog as hostile until proven otherwise.
Detection & Response
This is a technical threat. The detections below target the three most reliable observables: (1) execution of AI-branded installers from user-writable paths, (2) suspicious access to browser credential/session stores by non-browser processes, and (3) stealer-style staging in user profile directories.
Sigma Rules
---
title: AI-Branded Installer Executed from User-Writable Path
id: 3f8a1c92-7b4d-4e51-a9c6-2d1e5f7a8b9c
status: experimental
description: Detects execution of installers or executables masquerading as popular AI products launched from Downloads, Temp, or AppData — a consistent pattern in fake AI malware campaigns tracked by ANY.RUN (Apr-Oct 2026).
references:
- https://thehackernews.com/2026/10/fake-ai-attacks-spread-across-us.html
- https://attack.mitre.org/techniques/T1204/002/
author: Security Arsenal
date: 2026/10/20
tags:
- attack.execution
- attack.t1204.002
- attack.initial_access
- attack.t1189
logsource:
category: process_creation
product: windows
detection:
selection_path:
Image|contains:
- '\Downloads\'
- '\AppData\Local\Temp\'
- '\AppData\Roaming\'
- 'C:\Users\Public\'
selection_name:
Image|contains:
- 'chatgpt'
- 'chat-gpt'
- 'openai'
- 'claude'
- 'gemini'
- 'copilot'
- 'midjourney'
- 'perplexity'
- 'ai-assistant'
- 'ai_assistant'
- 'aisetup'
- 'ai-setup'
filter_legitimate:
# Microsoft Copilot and other sanctioned AI apps install under Program Files, not user-writable paths
Image|startswith:
- 'C:\Program Files\'
- 'C:\Program Files (x86)\'
condition: selection_path and selection_name and not filter_legitimate
falsepositives:
- Developers running locally built AI tooling from Downloads
- Portable AI utilities deliberately run from user directories (tune to your approved software list)
level: high
---
title: Non-Browser Process Accessing Browser Credential and Session Stores
id: 8c2e4d71-5f3a-4b96-c1d8-6e9a2f4b7d35
status: experimental
description: Detects non-browser processes reading Chrome/Edge/Brave credential databases, cookie stores, or DPAPI master keys — the core collection behavior of infostealers deployed via fake AI installer campaigns. Strong indicator of session/cookie theft used to bypass MFA.
references:
- https://thehackernews.com/2026/10/fake-ai-attacks-spread-across-us.html
- https://attack.mitre.org/techniques/T1555/003/
- https://attack.mitre.org/techniques/T1539/
author: Security Arsenal
date: 2026/10/20
tags:
- attack.credential_access
- attack.t1555.003
- attack.t1539
- attack.collection
logsource:
category: file_event
product: windows
detection:
selection_target:
TargetFilename|contains:
- '\Google\Chrome\User Data\'
- '\Microsoft\Edge\User Data\'
- '\BraveSoftware\Brave-Browser\User Data\'
selection_file:
TargetFilename|endswith:
- '\Login Data'
- '\Cookies'
- '\Local State'
- '\Web Data'
- '\Network\Cookies'
selection_dpapi:
TargetFilename|contains: '\Microsoft\Protect\'
filter_browsers:
Image|endswith:
- '\chrome.exe'
- '\msedge.exe'
- '\brave.exe'
- '\firefox.exe'
condition: (selection_target and selection_file) or selection_dpapi
# NOTE: Apply Image exclusion at your backend/pipeline level — combine with
# 'and not filter_browsers' if your file_event source populates Image.
falsepositives:
- Enterprise browser management/backup agents (whitelist by Image hash/path)
- EDR/DFIR collection tooling during sanctioned IR
level: critical
---
title: Suspicious Child Process from User-Run Installer (Stealer Staging)
id: b71d9e34-2c8f-4a57-9d63-1e4b6a8c5f02
status: experimental
description: Detects script interpreters and LOLBins spawned by MSIExec or installer executables running from user-writable paths — consistent with fake AI installers staging second-stage payloads and performing post-install theft while a decoy UI displays.
references:
- https://thehackernews.com/2026/10/fake-ai-attacks-spread-across-us.html
- https://attack.mitre.org/techniques/T1059/
author: Security Arsenal
date: 2026/10/20
tags:
- attack.execution
- attack.t1059
- attack.defense_evasion
- attack.t1218
logsource:
category: process_creation
product: windows
detection:
selection_parent:
ParentImage|contains:
- '\Downloads\'
- '\AppData\Local\Temp\'
- 'C:\Users\Public\'
selection_parent_type:
ParentImage|endswith:
- '\msiexec.exe'
- '\setup.exe'
- '\install.exe'
selection_child:
Image|endswith:
- '\powershell.exe'
- '\pwsh.exe'
- '\cmd.exe'
- '\wscript.exe'
- '\cscript.exe'
- '\mshta.exe'
- '\rundll32.exe'
- '\curl.exe'
- '\bitsadmin.exe'
- '\certutil.exe'
condition: selection_parent and selection_parent_type and selection_child
falsepositives:
- Legitimate installers spawning post-install scripts (rare for consumer-style apps; validate against approved software catalog)
level: high
Tuning guidance: Rule 2 is the highest-fidelity of the three — legitimate processes almost never read Login Data or Cookies outside the browser itself. In practice, deploy it with an explicit exclusion list for your managed browser backup/EDR tooling, and expect near-zero noise. Rule 1 will need a per-client approved-software list to stay quiet; start it in audit mode for 72 hours per tenant.
KQL — Microsoft Sentinel / Defender
// Hunt 1: AI-branded executables launched from user-writable paths (last 14 days)
DeviceProcessEvents
| where TimeGenerated > ago(14d)
| where FolderPath has_any ("\\Downloads\\", "\\AppData\\Local\\Temp\\", "\\AppData\\Roaming\\", "C:\\Users\\Public\\")
| where FileName has_any ("chatgpt","chat-gpt","openai","claude","gemini","copilot","midjourney","perplexity","ai-assistant","aisetup","ai-setup")
or ProcessCommandLine has_any ("chatgpt","openai","claude","gemini","midjourney","perplexity","ai setup","ai installer")
| where FolderPath !startswith "C:\\Program Files"
| summarize FirstSeen=min(TimeGenerated), LastSeen=max(TimeGenerated), Executions=count()
by DeviceName, AccountName, FileName, FolderPath, SHA256
| order by LastSeen desc;
// Hunt 2: Non-browser process accessing browser credential/cookie stores
DeviceFileEvents
| where TimeGenerated > ago(7d)
| where FolderPath has_any ("\\Google\\Chrome\\User Data\\", "\\Microsoft\\Edge\\User Data\\", "\\BraveSoftware\\")
| where FileName in~ ("Login Data","Cookies","Local State","Web Data")
| where InitiatingProcessFileName !in~ ("chrome.exe","msedge.exe","brave.exe","msedgewebview2.exe")
// Exclude your sanctioned EDR/backup agents here per tenant
| summarize AccessCount=count(), FirstSeen=min(TimeGenerated), LastSeen=max(TimeGenerated)
by DeviceName, InitiatingProcessFileName, InitiatingProcessFolderPath, InitiatingProcessSHA256
| order by AccessCount desc;
// Hunt 3: Network egress from freshly-dropped unsigned binaries in user paths
DeviceNetworkEvents
| where TimeGenerated > ago(7d)
| where InitiatingProcessFolderPath has_any ("\\Downloads\\", "\\AppData\\Local\\Temp\\", "\\AppData\\Roaming\\")
| where RemoteUrl !has_any ("microsoft.com","windowsupdate.com","google.com") // tune per tenant baseline
| join kind=inner (
DeviceFileEvents
| where TimeGenerated > ago(7d)
| where FolderPath has_any ("\\Downloads\\", "\\AppData\\Roaming\\")
| project DeviceName, FileName, SHA256, FileCreated=TimeGenerated
) on DeviceName, SHA256
| summarize Connections=count(), DistinctDestinations=dcount(RemoteUrl), FirstFile=min(FileCreated)
by DeviceName, InitiatingProcessFileName, SHA256
| where DistinctDestinations > 2
| order by Connections desc;
Hunt 2 is your bread-and-butter MSSP query — schedule it as an analytics rule at critical severity with a 4-hour lookback. Session theft is time-sensitive: every hour between cookie theft and session revocation is an hour the attacker has an authenticated session in your client's SaaS estate.
Velociraptor VQL
-- Hunt: Infostealer indicators — unsigned binaries in user paths + browser store access
-- Deploy as a client artifact across the fleet; triage hits by signer and hash reputation
LET unsigned_user_bins = SELECT
Pid,
Name,
Exe,
CommandLine,
Username,
Authenticode AS Signature,
CreateTime
FROM pslist()
WHERE (Exe =~ '(?i)Downloads|AppData\\(Local|Roaming)|Users\\Public'
AND Exe =~ '(?i)chatgpt|openai|claude|gemini|copilot|midjourney|perplexity|ai[-_ ]?(setup|assistant|install)')
LET browser_store_handles = SELECT
Pid,
Name,
Exe,
Username
FROM pslist()
WHERE Name !~ '(?i)chrome|msedge|brave|firefox'
AND CommandLine =~ '(?i)Login Data|Local State|Cookies|User Data'
SELECT * FROM unsigned_user_bins
UNION ALL
SELECT * FROM browser_store_handles
For deeper triage on a confirmed host, follow with a glob() collection of the browser stores themselves to preserve them for forensic timeline analysis before containment wipes them:
-- Artifact collection: preserve browser credential/session stores from a suspect host
SELECT FullPath, Size, Mtime, Atime
FROM glob(globs=[
'C:/Users/*/AppData/Local/Google/Chrome/User Data/*/Login Data',
'C:/Users/*/AppData/Local/Google/Chrome/User Data/*/Cookies',
'C:/Users/*/AppData/Local/Google/Chrome/User Data/Local State',
'C:/Users/*/AppData/Local/Microsoft/Edge/User Data/*/Login Data',
'C:/Users/*/AppData/Local/Microsoft/Edge/User Data/*/Cookies'
])
ORDER BY Mtime DESC
Unusually recent Mtime on Login Data or Cookies that doesn't correlate with user browser activity is a strong theft indicator.
Remediation & Hardening Script
The following PowerShell performs three functions on a suspect or at-risk Windows endpoint: (1) kills known-pattern stealer processes, (2) enumerates AI-branded executables in user-writable paths for triage, and (3) applies WDAC/AppLocker-style hardening via SmartScreen and browser policy to reduce repeat exposure. Run in an elevated context; adapt paths per tenant.
#Requires -RunAsAdministrator
# Fake-AI infostealer triage + hardening — Security Arsenal IR kit
$aiPattern = 'chatgpt|chat-gpt|openai|claude|gemini|copilot|midjourney|perplexity|ai[-_ ]?(setup|assistant|install)'
$userPaths = @("$env:SystemDrive\Users\*\Downloads", "$env:SystemDrive\Users\*\AppData\Roaming", "$env:SystemDrive\Users\Public")
# 1) Enumerate and kill running AI-branded processes NOT under Program Files
Get-Process | Where-Object {
$_.Path -and ($_.Path -notlike "$env:ProgramFiles*") -and
($_.Path -notlike "${env:ProgramFiles(x86)}*") -and
($_.Name -match $aiPattern)
} | ForEach-Object {
Write-Warning "Suspicious process: $($_.Name) [$($_.Id)] at $($_.Path)"
Stop-Process -Id $_.Id -Force
}
# 2) Inventory AI-branded binaries in user-writable paths for IR triage
$triage = foreach ($p in $userPaths) {
Get-ChildItem -Path $p -Recurse -Include *.exe,*.msi -ErrorAction SilentlyContinue |
Where-Object { $_.Name -match $aiPattern } |
Select-Object FullName, Length, CreationTime,
@{N='SHA256';E={(Get-FileHash $_.FullName -Algorithm SHA256).Hash}},
@{N='Signer';E={(Get-AuthenticodeSignature $_.FullName).SignerCertificate.Subject}}
}
$triage | Export-Csv -Path "C:\IR-Triage-FakeAI-$(Get-Date -Format 'yyyyMMdd-HHmm').csv" -NoTypeInformation
$triage | Format-Table -AutoSize
# 3) Hardening: enforce SmartScreen (blocks known-malicious downloads) and browser cookie encryption policy
Set-ItemProperty -Path 'HKLM:\SOFTWARE\Policies\Microsoft\Windows\System' -Name 'EnableSmartScreen' -Value 1 -Force -ErrorAction SilentlyContinue
New-Item -Path 'HKLM:\SOFTWARE\Policies\Microsoft\Edge' -Force | Out-Null
Set-ItemProperty -Path 'HKLM:\SOFTWARE\Policies\Microsoft\Edge' -Name 'SmartScreenEnabled' -Value 1 -Force
New-Item -Path 'HKLM:\SOFTWARE\Policies\Google\Chrome' -Force | Out-Null
Set-ItemProperty -Path 'HKLM:\SOFTWARE\Policies\Google\Chrome' -Name 'SafeBrowsingProtectionLevel' -Value 2 -Force
# 4) Audit: list outbound connections from processes in user paths (stealer exfil check)
Get-NetTCPConnection -State Established -ErrorAction SilentlyContinue |
ForEach-Object {
$proc = Get-Process -Id $_.OwningProcess -ErrorAction SilentlyContinue
if ($proc.Path -and $proc.Path -match 'Downloads|AppData\\Roaming|Users\\Public') {
[PSCustomObject]@{ Process=$proc.Name; Path=$proc.Path; Remote=$_.RemoteAddress; Port=$_.RemotePort }
}
} | Sort-Object -Unique Process, Remote
Write-Host "`n[+] Triage complete. CRITICAL NEXT STEPS (manual):" -ForegroundColor Yellow
Write-Host " - Revoke ALL active sessions in M365/Entra ID and Google Workspace for affected users"
Write-Host " - Force password reset + re-enroll MFA (session theft defeats existing tokens)"
Write-Host " - Clear browser cookies/sessions on the endpoint AFTER forensic collection"
Write-Host " - Submit hashes from triage CSV to your sandbox (e.g., ANY.RUN) for family attribution"
Remediation
There is no patch for this threat — remediation is containment, credential hygiene, and preventive control. Execute in this order:
- Contain the host immediately. Network-isolate any endpoint where a fake AI installer executed. Stealers exfiltrate in seconds, so assume compromise occurred.
- Revoke sessions first, reset passwords second. Stolen cookies bypass password resets. In Entra ID: revoke sign-in sessions and refresh tokens for the affected user; in Google Workspace: sign out all sessions and revoke OAuth grants. Then force credential reset and MFA re-registration. Review Conditional Access sign-in logs for impossible-travel or token-replay anomalies in the preceding 72 hours.
- Enable token protection where available. Microsoft Entra token protection (Conditional Access session controls) cryptographically binds tokens to the device, blunting cookie replay. Chrome's App-Bound Encryption (enforced by default in current Chrome builds on Windows) materially raises the bar for cookie theft — verify it hasn't been disabled by policy.
- Establish a sanctioned AI software catalog. Publish an approved list of AI tools with canonical download sources; enforce via AppLocker/WDAC or your EDR's application control. Block execution of unsigned installers from user-writable paths as policy, not as an alert.
- Block the delivery vector. DNS-filter known lookalike/typosquat domains, enable Safe Browsing Enhanced Protection / SmartScreen enterprise-wide, and consider blocking search-engine ads at the proxy for software-download queries.
- Hunt retroactively. Run the KQL hunts above across a 6-month window (matching the ANY.RUN campaign timeframe) to catch historical executions you missed.
- User communication. Notify affected users that their saved browser credentials across ALL sites (personal and business) should be considered compromised — stealer logs get sold wholesale.
For MSSPs: the operational pressure noted in the reporting is real — these are high-frequency, low-severity-per-event alerts with occasional catastrophic outcomes. Automate the session-revocation step via SOAR so analyst response time doesn't become the breach window.
Related Resources
Security Arsenal Incident Response Services AlertMonitor Platform Book a SOC Assessment incident-response Intel Hub
Is your security operations ready?
Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.