A single OTX pulse from AlienVault, sourced from Zscaler ThreatLabz research, documents one of the most operationally significant campaigns of late 2026: a Chinese-speaking threat actor conducting widespread AI-assisted cyberattacks at global scale by chaining the ARTEX penetration testing framework with multiple commercial AI coding assistants — Claude Code, Codex, and Grok.
Initially surfaced as targeted intrusions against South Korean financial institutions in October 2026, follow-on analysis revealed the campaign's true breadth: victims spanning finance, retail, education, technology, telecommunications, hospitality, and government across at least 19 countries, including the United States, United Kingdom, Canada, India, Japan, UAE, Nigeria, Kenya, and South Africa.
The attack chain represents an evolution in offensive tradecraft rather than a novel malware family. The operator uses large language models as an orchestration layer — tasking AI coding assistants to generate reconnaissance scripts, weaponize the ARTEX framework's exploitation and post-exploitation modules, automate target enumeration, and iterate on attack tooling in real time. This multi-model orchestration dramatically compresses the time between initial access and objective completion, and allows a single operator to run concurrent operations against dozens of organizations that would previously have required a full red team.
Objectives observed align with financial crime and data theft: harvesting credentials and session data from financial sector targets, exfiltrating proprietary and customer data from technology and retail victims, and establishing durable access in government and telecommunications environments consistent with longer-term intelligence collection.
Threat Actor / Malware Profile
ARTEX Framework
ARTEX is a modular penetration testing and post-exploitation framework being abused as the operational backbone of this campaign. Key characteristics relevant to defenders:
- Distribution method: Initial access is achieved through automated reconnaissance and exploitation of internet-facing services — vulnerable web applications, exposed remote access services, and phished credentials validated against VPN and SSO portals. The AI orchestration layer generates and tests exploit variants against enumerated targets at machine speed.
- Payload behavior: ARTEX deploys staged post-exploitation modules for host enumeration, credential harvesting (LSASS access, browser credential stores, cached tokens), lateral movement via SMB/WinRM/RDP, and data staging. AI-generated helper scripts are frequently dropped as obfuscated Python or PowerShell one-offs — each sample unique, defeating hash-based detection.
- C2 communication: Framework C2 traffic is proxied through legitimate cloud infrastructure and content delivery networks, blending with normal SaaS egress. Because AI assistants regenerate tooling per-target, C2 endpoints rotate frequently; infrastructure-level indicators have a short shelf life.
- Persistence mechanism: Observed persistence includes scheduled tasks masquerading as system maintenance jobs, run-key registry entries, and webshells planted on compromised web-facing servers for re-entry.
- Anti-analysis techniques: Per-victim polymorphism is the defining trait. Every engagement features freshly generated scripts with randomized variable names, junk code, and unique packing — a direct consequence of AI-assisted tooling generation. Traditional signature-based AV and IOC matching are largely ineffective; behavioral detection is mandatory.
The AI Orchestration Layer
The actor's use of Claude Code, Codex, and Grok as force multipliers means defenders should expect: rapid tool iteration after detection, near-zero dwell time between reconnaissance and exploitation, and attack tooling that does not match any known malware family's signatures. Detection engineering must focus on TTPs — not artifacts.
IOC Analysis
The pulse's indicator set is restricted at the API level — no IPs, domains, hashes, or URLs were distributed with this pulse. This is consistent with a campaign whose indicators are per-victim and short-lived. SOC teams should operationalize this intelligence as follows:
- Do not wait for IOC feeds. Pull the full indicator set directly from the Zscaler research reference and the original OTX pulse via authenticated API access, and ingest into your SIEM/TIP with a 72-hour TTL given the rotation speed.
- Pivot to behavioral indicators: ARTEX module execution patterns, anomalous scripting engine usage (PowerShell, Python) spawned from non-standard parent processes, and enumeration command sequences executed in rapid succession (a hallmark of AI-generated automation).
- Tooling: Decode and detonate any recovered scripts in a sandbox (ANY.RUN, Joe Sandbox, or a local CAPE instance). Use
oletools,peframe, and CyberChef for script deobfuscation. For AI-generated PowerShell, AMSI telemetry and Script Block Logging (Event ID 4104) are your highest-fidelity data sources. - Network-side: Baseline egress to AI service domains and developer tooling endpoints. Unexpected workstation-level traffic to LLM APIs, or AI-assistant CLI artifacts on servers, is itself a detection opportunity.
Detection Engineering
---
title: ARTEX Framework Post-Exploitation - Rapid Automated Enumeration
id: 7f3a2b11-9c4e-4d21-a8f6-1e5b9c02aa01
status: experimental
description: Detects rapid sequences of host/network enumeration commands characteristic of AI-orchestrated automated reconnaissance used with the ARTEX framework. Burst execution of multiple enumeration utilities within a short window indicates scripted post-exploitation.
references:
- https://www.zscaler.com/blogs/security-research/inside-global-ai-assisted-hacking-campaign
author: Security Arsenal Threat Intelligence
date: 2026/11/22
logsource:
category: process_creation
product: windows
detection:
selection_img:
Image|endswith:
- '\ipconfig.exe'
- '\net.exe'
- '\nltest.exe'
- '\arp.exe'
- '\systeminfo.exe'
- '\whoami.exe'
- '\quser.exe'
- '\route.exe'
selection_args:
CommandLine|contains:
- ' /all'
- ' user'
- ' group'
- ' localgroup'
- ' /domain'
- 'dclist'
- ' /query'
condition: selection_img and selection_args
timeframe: 5m
falsepositives:
- Legitimate administrative scripts and IT inventory tooling
- Login scripts executing enumeration for drive mapping
level: medium
tags:
- attack.discovery
- attack.t1033
- attack.t1018
- attack.t1016
---
title: Suspicious AI-Generated Script Execution - Encoded PowerShell From Office or Browser Parent
id: 8a1c4d22-3b7f-4e19-b2d4-9f8a1c55bb02
status: experimental
description: Detects PowerShell launched with encoded or obfuscated commands from browser or Office parent processes, consistent with AI-generated one-off loader scripts dropped during ARTEX campaign initial access.
references:
- https://www.zscaler.com/blogs/security-research/inside-global-ai-assisted-hacking-campaign
author: Security Arsenal Threat Intelligence
date: 2026/11/22
logsource:
category: process_creation
product: windows
detection:
selection_child:
Image|endswith:
- '\powershell.exe'
- '\pwsh.exe'
selection_parent:
ParentImage|endswith:
- '\winword.exe'
- '\excel.exe'
- '\outlook.exe'
- '\chrome.exe'
- '\msedge.exe'
- '\firefox.exe'
- '\rundll32.exe'
selection_flags:
CommandLine|contains:
- ' -enc'
- ' -e '
- 'FromBase64String'
- 'IEX'
- 'Invoke-Expression'
- 'DownloadString'
- 'bypass'
- '-w hidden'
condition: all of selection_*
falsepositives:
- Rare; some enterprise management tools use encoded commands, but not typically from browser or Office parents
level: high
tags:
- attack.execution
- attack.t1059.001
- attack.t1027
---
title: ARTEX Campaign Persistence - Scheduled Task Masquerading As System Maintenance
id: 2b9e5f33-6d1a-4c38-c7e5-2a4b6d88cc03
status: experimental
description: Detects creation of scheduled tasks with names imitating legitimate system maintenance that execute scripting engines or binaries from user-writable paths, matching persistence behavior observed in the AI-assisted ARTEX campaign.
references:
- https://www.zscaler.com/blogs/security-research/inside-global-ai-assisted-hacking-campaign
author: Security Arsenal Threat Intelligence
date: 2026/11/22
logsource:
category: process_creation
product: windows
detection:
selection_schtasks:
Image|endswith: '\schtasks.exe'
CommandLine|contains: '/create'
selection_path:
CommandLine|contains:
- '\AppData\'
- '\Temp\'
- '\ProgramData\'
- '\Users\Public\'
selection_engine:
CommandLine|contains:
- 'powershell'
- 'pwsh'
- 'python'
- 'wscript'
- 'mshta'
- 'rundll32'
condition: all of selection_*
falsepositives:
- Software updaters and enterprise deployment tools registering tasks in ProgramData
level: high
tags:
- attack.persistence
- attack.t1053.005
- attack.execution
// Hunt: AI-orchestrated ARTEX campaign behaviors — burst enumeration + suspicious script execution + anomalous egress
// Microsoft Sentinel / MDE
let lookback = 7d;
let enumProcs = dynamic(["ipconfig.exe","net.exe","nltest.exe","arp.exe","systeminfo.exe","whoami.exe","quser.exe","route.exe","nslookup.exe"]);
let enumBurst = DeviceProcessEvents
| where TimeGenerated > ago(lookback)
| where FileName in~ (enumProcs)
| summarize EnumCount = count(), Commands = make_set(strcat(FileName, " ", ProcessCommandLine), 20) by DeviceId, DeviceName, AccountName, bin(TimeGenerated, 5m)
| where EnumCount >= 5;
let suspiciousScripts = DeviceProcessEvents
| where TimeGenerated > ago(lookback)
| where FileName in~ ("powershell.exe","pwsh.exe","python.exe","wscript.exe","mshta.exe")
| where ProcessCommandLine has_any ("-enc","FromBase64String","IEX","Invoke-Expression","DownloadString","bypass","-w hidden")
| project ScriptTime=TimeGenerated, DeviceId, DeviceName, InitiatingProcessFileName, ScriptCmd=ProcessCommandLine;
let unusualEgress = DeviceNetworkEvents
| where TimeGenerated > ago(lookback)
| where RemoteUrl has_any ("claude.ai","api.anthropic.com","openai.com","x.ai","grok") or RemoteUrl endswith ".workers.dev" or RemoteUrl endswith ".trycloudflare.com"
| where InitiatingProcessFileName !in~ ("chrome.exe","msedge.exe","firefox.exe","Code.exe")
| summarize NetEvents=count(), Targets=make_set(RemoteUrl, 10) by DeviceId, InitiatingProcessFileName, bin(TimeGenerated, 1h);
enumBurst
| join kind=leftouter suspiciousScripts on DeviceId, DeviceName
| join kind=leftouter unusualEgress on DeviceId
| project TimeGenerated, DeviceName, AccountName, EnumCount, Commands, ScriptTime, InitiatingProcessFileName, ScriptCmd, Targets
| order by TimeGenerated desc;
# ARTEX AI-Assisted Campaign - Endpoint Hunt Script
# Checks for persistence artifacts, suspicious script execution logs, and anomalous outbound connections
# Run elevated on suspect hosts; review output for follow-up triage
Write-Output "=== [1] Scheduled Tasks Executing From User-Writable Paths ==="
Get-ScheduledTask | ForEach-Object {
$task = $_
$_.Actions | Where-Object {
$_.Execute -match 'AppData|Temp|ProgramData|Users\\Public' -or
$_.Arguments -match 'AppData|Temp|ProgramData|Users\\Public'
} | ForEach-Object {
[PSCustomObject]@{
TaskName = $task.TaskName
TaskPath = $task.TaskPath
Execute = $_.Execute
Arguments = $_.Arguments
Author = $task.Author
}
}
} | Format-Table -AutoSize
Write-Output "=== [2] Run-Key Persistence In User Hives ==="
$runKeys = @(
'HKCU:\Software\Microsoft\Windows\CurrentVersion\Run',
'HKCU:\Software\Microsoft\Windows\CurrentVersion\RunOnce',
'HKLM:\Software\Microsoft\Windows\CurrentVersion\Run',
'HKLM:\Software\Microsoft\Windows\CurrentVersion\RunOnce'
)
foreach ($key in $runKeys) {
if (Test-Path $key) {
Get-ItemProperty -Path $key | ForEach-Object {
$_.PSObject.Properties | Where-Object { $_.Name -notmatch '^PS' -and $_.Value -match 'AppData|Temp|powershell|python|wscript|mshta' } |
ForEach-Object { Write-Output ("{0} :: {1} = {2}" -f $key, $_.Name, $_.Value) }
}
}
}
Write-Output "=== [3] PowerShell Script Block Logs - Obfuscation Indicators (Last 7 Days) ==="
Get-WinEvent -FilterHashtable @{LogName='Microsoft-Windows-PowerShell/Operational'; Id=4104; StartTime=(Get-Date).AddDays(-7)} -ErrorAction SilentlyContinue |
Where-Object { $_.Message -match 'FromBase64String|Invoke-Expression|DownloadString|Net\.WebClient|Add-MpPreference' } |
Select-Object -First 50 TimeCreated, @{N='Snippet';E={$_.Message.Substring(0,[Math]::Min(300,$_.Message.Length))}} |
Format-List
Write-Output "=== [4] Active Connections To AI Service / Cloudflare Tunnel Infrastructure ==="
$aiPatterns = 'anthropic|openai|x\.ai|claude|grok|workers\.dev|trycloudflare'
Get-NetTCPConnection -State Established -ErrorAction SilentlyContinue | ForEach-Object {
$conn = $_
try {
$proc = Get-Process -Id $conn.OwningProcess -ErrorAction Stop
$dns = Resolve-DnsName -Name $conn.RemoteAddress -ErrorAction SilentlyContinue | Select-Object -First 1
if ($proc.ProcessName -notmatch 'chrome|msedge|firefox') {
[PSCustomObject]@{
Process = $proc.ProcessName
PID = $conn.OwningProcess
RemoteIP = $conn.RemoteAddress
RemotePort = $conn.RemotePort
ResolvedName = $dns.NameHost
}
}
} catch {}
} | Where-Object { $_.ResolvedName -match $aiPatterns -or $_.Process -match 'powershell|python|pwsh' } | Format-Table -AutoSize
Write-Output "=== [5] Recently Created Executables In Staging Directories (Last 72h) ==="
$stagingPaths = @("$env:TEMP","$env:LOCALAPPDATA\Temp","$env:ProgramData","C:\Users\Public")
foreach ($path in $stagingPaths) {
if (Test-Path $path) {
Get-ChildItem -Path $path -Recurse -Include *.exe,*.ps1,*.py,*.dll -ErrorAction SilentlyContinue |
Where-Object { $_.CreationTime -gt (Get-Date).AddHours(-72) } |
Select-Object FullName, CreationTime, Length
}
}
Response Priorities
Immediate (0-4 hours)
- Ingest the full indicator set from the referenced Zscaler research and OTX pulse into blocklists; apply a 72-hour TTL given infrastructure rotation speed.
- Deploy the Sigma rules and KQL hunt above; prioritize alerts showing enumeration bursts correlating with suspicious script execution on the same host.
- Audit outbound traffic to LLM API endpoints from servers and non-developer workstations — developer tooling on production servers is a high-fidelity anomaly in this campaign's context.
- Isolate any host exhibiting the enumeration-burst-plus-encoded-script pattern pending forensic triage.
24 Hours
- Because credential harvesting and data theft are confirmed campaign objectives, force credential rotation for any account that authenticated from or to a suspected host, with priority on privileged and financial-system accounts.
- Invalidate active sessions and refresh tokens (especially cloud SSO sessions) for impacted users; review MFA enrollment and anomalous consent grants.
- Review LSASS access events (Sysmon Event ID 10) and DCSync-style replication requests for evidence of credential theft.
- Verify email and financial-process integrity at finance-sector business units — check for mailbox rules, payment instruction changes, and BEC pre-positioning.
1 Week
- Harden internet-facing services: patch exploited web application classes, enforce phishing-resistant MFA on VPN/SSO, and reduce exposed attack surface the automated reconnaissance phase can enumerate.
- Deploy or tighten PowerShell Script Block Logging, AMSI integration, and constrained language mode on servers; block unsigned script execution via WDAC/AppLocker where feasible.
- Establish an egress policy governing AI-service and developer-tooling domains; alert on first-seen connections from server VLANs.
- Conduct purple-team validation of the enumeration-burst and persistence detections, and brief threat hunting teams on AI-generated tooling polymorphism — train analysts to hunt behavior, not hashes.
Related Resources
Security Arsenal Incident Response Managed SOC & MDR Services AlertMonitor Threat Detection From The Dark Side Intel Hub
Is your security operations ready?
Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.