Skip to navigation

Threat reportMalwareTL-2026-4072

Fake AI Tool Campaigns Targeting Enterprises (ChatGPT, Claude, Copilot, Gemini, GitHub Copilot CLI lures)

mediumACTIVE

ANY.RUN analysis of 1,258 public sandbox submissions (Apr 15 - Oct 5, 2026) ties a family of campaigns to counterfeit AI sites, fake CLI installers (InstallFix), ClickFix command-paste lures, fake sign-in windows and trojanized desktop apps that steal browser credentials, cookies and session tokens. First disclosed .

Key facts

IdentifiersTL-2026-4072
Also known asInstallFix AI-brand impersonation, Fake AI tool campaigns
SeverityMEDIUM
First disclosed
Affected productsMicrosoft Windows (endpoints running user-installed AI tools); Apple macOS (Terminal command-paste lures); Anthropic / OpenAI / Microsoft / Google Brands impersonated: Claude, Claude Code, ChatGPT, Codex, Copilot, GitHub Copilot CLI, Gemini (no vendor product vulnerability)
MalwareAMOS, Amatera Stealer, Lumma Stealer - S1213, MacSync Infostealer, Noodlophile Stealer
Toolsmshta.exe
MITRE ATT&CK techniquesT1027 Obfuscated Files or Information, T1036 Masquerading, T1056.002 GUI Input Capture, T1059.003 Windows Command Shell, T1059.004 Unix Shell, T1140 Deobfuscate/Decode Files or Information, T1204.002 Malicious File, T1204.004 Malicious Copy and Paste, T1218.005 Mshta, T1539 Steal Web Session Cookie, T1555.003 Credentials from Web Browsers, T1583.006 Web Services and 1 more in the attack flow
MotivationUNKNOWN
Target sectorstechnology, telecoms, education, consulting, manufacturing, health, government administration, finance
Target regionsindia, united states of america, Global
Detection rules9 (1 shown free)
Sources7 cited sources

Detection coverage

As of 2026-10-21, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-4072 across Splunk SPL, Microsoft KQL and Sigma, covering 21 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: mshta.exe Fetching Remote Payload from Command Shell - Fake AI Tool InstallFix (TL-2026-4072)

Detects the Windows InstallFix/ClickFix chain documented by Push Security and Malwarebytes for fake Claude Code install pages: cmd.exe, PowerShell or Explorer spawns mshta.exe with a remote URL or script protocol to run the Amatera infostealer payload. Behaviour-first: fires on the process chain regardless of domain, and flags command lines containing known research IOC strings.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1218.005, T1059.003, T1204.004
  • Data sources: sysmon, windows_process_creation
  • Severity: high
  • Confidence: high

Splunk SPL

index=sysmon sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" EventCode=1 Image="*mshta.exe" earliest=-24h
| where match(Image,"(?i)[^a-z0-9]mshta[.]exe$")
| where match(CommandLine,"(?i)(https?://|vbscript:|javascript:)")
| where match(ParentImage,"(?i)[^a-z0-9](cmd|powershell|pwsh|explorer|wscript|cscript)[.]exe$")
| eval ioc_hit=if(match(CommandLine,"(?i)(verification-claude-cdn|claud-code|claude-code|update-version|a2abotnet|download-version)"),"research_ioc","behavior_only")
| stats count min(_time) as firstTime max(_time) as lastTime values(CommandLine) as CommandLine values(ParentCommandLine) as ParentCommandLine values(ioc_hit) as ioc_hit by host User ParentImage Image
| convert ctime(firstTime) ctime(lastTime)

Microsoft KQL

DeviceProcessEvents
| where Timestamp > ago(24h)
| where FileName =~ 'mshta.exe'
| where ProcessCommandLine matches regex @'(?i)(https?://|vbscript:|javascript:)'
| where InitiatingProcessFileName in~ ('cmd.exe','powershell.exe','pwsh.exe','explorer.exe','wscript.exe','cscript.exe')
| extend ioc_hit = iff(ProcessCommandLine matches regex @'(?i)(verification-claude-cdn|claud-code|claude-code|update-version|a2abotnet|download-version)', 'research_ioc', 'behavior_only')
| project Timestamp, DeviceName, AccountName, InitiatingProcessFileName, InitiatingProcessCommandLine, FileName, ProcessCommandLine, ioc_hit, SHA256

Sigma

title: Mshta Executing Remote Payload From Command Shell - Fake AI Tool InstallFix
id: 1b7e4c9a-6d23-4e58-a0f1-3c5d7e9b1a04
status: experimental
description: Detects mshta.exe launched by a command shell, Explorer or script host with a remote URL or script protocol, the Windows InstallFix/ClickFix pattern used by fake Claude Code and AI tool install pages to deliver Amatera stealer (TL-2026-4072).
references:
    - https://www.malwarebytes.com/blog/news/2026/03/fake-claude-code-install-pages-hit-windows-and-mac-users-with-infostealers
    - https://www.helpnetsecurity.com/2026/03/09/fake-claude-code-install-pages-installfix-attacks/
author: Threadlinqs Team
date: 2026-10-21
tags:
    - attack.defense-evasion
    - attack.t1218.005
    - attack.execution
    - attack.t1059.003
    - attack.t1204.004
logsource:
    category: process_creation
    product: windows
detection:
    selection_img:
        Image|endswith: '\mshta.exe'
    selection_remote:
        CommandLine|contains:
            - 'http://'
            - 'https://'
            - 'vbscript:'
            - 'javascript:'
    selection_parent:
        ParentImage|endswith:
            - '\cmd.exe'
            - '\powershell.exe'
            - '\pwsh.exe'
            - '\explorer.exe'
            - '\wscript.exe'
            - '\cscript.exe'
    condition: all of selection_*
falsepositives:
    - Legacy internal HTA applications launched from a script with an intranet URL
    - Software deployment tooling that still uses mshta against internal servers
level: high

Known false positives:

  • Legacy internal HTA applications launched with an intranet URL
  • Software deployment tooling that still invokes mshta against internal servers

Tuning notes: Adjust the index name to where Sysmon EventCode 1 is stored. Allow-list internal HTA hosts by excluding CommandLine values that match your intranet domains. Where mshta is not needed, enforce application control and treat any hit as high priority.

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)
  • ClickFix Paste-and-Run Command Launched From Explorer or Run Dialog · unlock with TL Creds
  • Browser Credential and Cookie Store Files Created or Copied by Non-Browser Process · unlock with TL Creds
  • LOLBin or Script Host Network Activity to Fake AI Infrastructure and Abused Hosting · unlock with TL Creds
  • Download Utility Fetching From Free Hosting With AI Brand Keywords · unlock with TL Creds
  • macOS Curl-to-Shell Install One-Liner From Non-Official AI Brand URL · unlock with TL Creds
  • macOS Fake Password Prompt and Keychain Access by Infostealer · unlock with TL Creds
  • macOS Terminal Base64-Decoded Second Stage Piped to Shell - Fake AI Install Lure · unlock with TL Creds
  • Web Proxy Access to Fake AI Brand Install Infrastructure and Lookalike Hosting · unlock with TL Creds

Indicators of compromise

As of 2026-10-21, Threadlinqs Intelligence tracks 21 indicators of compromise for TL-2026-4072.

Sample indicators (3 of 21)

  • Domain: download-version[.]1-9-18[.]com
  • Domain: perplexity-ai[.]online
  • Domain: code[.]verification-claude-cdn[.]beer

Defanged, TLP:CLEAR. Full set of 21 indicators: unlock with TL Creds.

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

Technical analysis

Claude is the most impersonated brand, followed by ChatGPT/Codex, and the technology sector was hit in all seven tracked campaigns.

ANY.RUN, as relayed by Cyber Security News (2026-10-21) and several secondary outlets, examined 1,258 public sandbox submissions recorded between 2026-04-15 and 2026-10-05 and grouped them into seven campaigns that abuse the popularity of AI tools. Reported lure brands are Claude (about 60% of analyses; 601 records in one secondary tally), ChatGPT/Codex (about 33%; 418 records), Microsoft Copilot, GitHub Copilot CLI, Google Gemini and Perplexity. Activity peaked in May 2026 (358 analyses) and ran at 108-151 analyses per month from June to September.

The common trait is that the employee performs the decisive step. Lures arrive through sponsored search listings, SEO poisoning, malicious ads and public AI conversations, and lead to fraudulent landing pages hosted on GitHub Pages, GitLab Pages, Cloudflare Pages and Google Sites. Reported delivery patterns are: fake CLI installers (InstallFix, about 27% of analyses) in which a cloned installation page swaps the legitimate one-liner for a malicious one; classic ClickFix flows that begin with a fake download page (about 17%); trojanized desktop installers that run a legitimate AI application while collecting browser data; and fake sign-in windows (for example a fraudulent Google sign-in presented in fake 'Claude Ads' lures aimed at marketing staff). Terminal command execution accounted for 341 instances and rogue download sites for 220 in one secondary breakdown. Developers are lured with fake GitHub Copilot CLI install pages that put source code, cloud credentials and AI API keys at risk.

On Windows, the InstallFix command documented by Push Security and Malwarebytes spawns cmd.exe and invokes mshta.exe against a remote URL to run the payload (Amatera infostealer). On macOS, the pasted one-liner pulls a base64-obfuscated second-stage script that downloads a binary, strips file attributes (quarantine) and sets it executable. ANY.RUN also observed in early June 2026 a ClickFix campaign mimicking Claude Code and OpenAI Codex with pages on sites.google.com that uses hidden characters in commands so a second command runs in the foreground and downloads an infostealer targeting API keys, browser-saved passwords and crypto wallets. Stolen passwords, cookies and authentication tokens enable session hijacking and access to cloud dashboards, ad accounts and billing without needing the password.

Top affected sectors in the ANY.RUN data: technology (7 of 7 campaigns), telecommunications (6), education (5), consulting (5), manufacturing, healthcare, government and finance (4 each). Country-tagged submissions came mainly from India (about 20%) and the United States (about 16%). The primary ANY.RUN-derived article names no malware families, IOCs, CVEs or actors. The network IOCs and malware families recorded here come from separate vendor reporting on AI-brand impersonation (Sophos, Cloud Security Alliance, Push Security/Malwarebytes) and are related-campaign indicators, not confirmed members of the seven ANY.RUN clusters. No attribution is made.

MITRE ATT&CK techniques by tactic:

Defense Evasion

T1027 Obfuscated Files or Information; T1036 Masquerading; T1140 Deobfuscate/Decode Files or Information; T1218.005 Mshta

Credential Access

T1056.002 GUI Input Capture; T1539 Steal Web Session Cookie; T1555.003 Credentials from Web Browsers

Execution

T1059.003 Windows Command Shell; T1059.004 Unix Shell; T1204.002 Malicious File; T1204.004 Malicious Copy and Paste

Resource Development

T1583.006 Web Services; T1583.008 Malvertising

Remediation

Immediate actions

  • Block or alert on egress to the related domains listed in the IOC set and hunt proxy/DNS logs for them
  • Alert on cmd.exe spawning mshta.exe with a remote URL, and on terminal one-liners that decode base64 and pipe to a shell
  • Revoke and rotate browser-saved credentials, session cookies, cloud and AI API keys on any device that ran a pasted install command or untrusted AI installer

Workarounds

  • Restrict mshta.exe execution via application control where not required
  • Block sponsored-result and newly registered lookalike domains for AI brands at the secure web gateway

Longer-term hardening

  • Allow-list approved AI tools and distribute them through managed software channels rather than search-engine results
  • Train staff, developers and marketing users that official AI CLI install commands must be copied only from vendor-owned domains, and never to paste commands from web pages into a terminal or Run dialog without review
  • Enforce phishing-resistant MFA and short session lifetimes with token-binding or conditional access to limit stolen-cookie replay
  • Monitor ad accounts and billing access for anomalous sign-ins

Affected and fixed versions

  • Microsoft — Windows (endpoints running user-installed AI tools)
    Vulnerable versions: user-execution based; not version specific
  • Apple — macOS (Terminal command-paste lures)
    Vulnerable versions: user-execution based; not version specific
  • Anthropic / OpenAI / Microsoft / Google — Brands impersonated: Claude, Claude Code, ChatGPT, Codex, Copilot, GitHub Copilot CLI, Gemini (no vendor product vulnerability)

Timeline

  • Push Security research on InstallFix fake Claude Code install pages delivering Amatera stealer is reported by Help Net Security; Malwarebytes covers the same campaign the same month
  • Start of the ANY.RUN observation window of 1,258 public sandbox analyses linked to fake AI tool campaigns
  • Activity peaks in May 2026 with 358 analyses (date set to month end; source gives month only); June-September then runs at 108-151 analyses per month
  • In early June (month-level date) ANY.RUN spots a ClickFix campaign mimicking Claude Code and OpenAI Codex on sites.google.com that uses hidden characters to run a second command and download an infostealer
  • End of the ANY.RUN observation window; seven campaigns identified, with technology targeted in all seven
  • Cyber Security News publishes the ANY.RUN findings on fake AI tools putting enterprises at risk

Sources cited

Previous and next in malware

Drafted by the Threadlinqs research pipeline from 7 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