Security teams have spent years conditioning users to trust well-known domains: microsoft.com, google.com, github.com. Attackers learned long ago that hosting phishing content on legitimate infrastructure defeats both URL filtering and user intuition. According to recent reporting from Huntress, that same playbook has now fully matured against AI platforms — and it's working.
Threat actors are abusing Claude Artifacts (shareable, Claude-hosted interactive content), shared AI conversation links, sponsored search results, and ClickFix-style fake error lures to push malware to users who believe they're interacting with a legitimate Anthropic-hosted resource. The malicious content lives on infrastructure users have been trained to trust, which neutralizes a significant portion of traditional web filtering and user-awareness defenses.
If your users are adopting AI assistants — and they are, whether sanctioned or not — this is an initial-access vector your SOC needs coverage on today.
Technical Analysis: How the Attack Chain Works
This is not a vulnerability in Anthropic's platform in the traditional CVE sense. There is no patchable bug. The abuse vector is a feature: the ability to create and publicly share AI-generated content under a trusted domain.
Delivery Vectors Observed
-
Weaponized Claude Artifacts. Attackers generate interactive Artifacts — HTML/JavaScript applications hosted on Claude's sharing infrastructure — and distribute links via phishing, SEO poisoning, or malvertising. Because the URL resolves to a legitimate, reputable AI domain, it sails through category-based web filters and looks benign in email gateways.
-
Shared AI conversations as lure pages. Shared conversation links are being used to host convincing, attacker-scripted dialogues — for example, a fake "AI support session" that walks the victim through a ClickFix execution. The conversational format lends false credibility: users see what looks like a helpful assistant instructing them to paste a command.
-
Sponsored search poisoning. Attackers purchase sponsored results on search engines targeting AI-related queries ("Claude download," "AI assistant installer," "ChatGPT desktop app"). The ads lead to lookalike pages or directly to malicious shared content.
-
ClickFix execution. The payload stage follows the now-standard ClickFix pattern: the page presents a fake error, CAPTCHA, or "verification" step that instructs the user to open the Run dialog (Win+R) or Terminal and paste a command from the clipboard. The clipboard has been silently populated with a malicious one-liner — typically a PowerShell or
mshta/curl-based command that downloads and executes a second-stage payload.
Why This Defeats Existing Controls
- URL reputation is useless when the hosting domain is a major AI provider with a pristine reputation score.
- Email gateway rewriting/sandboxing sees a clean HTTPS link to a trusted SaaS property.
- User training tells people to look for suspicious domains — this campaign has none.
- The ClickFix technique specifically bypasses browser download controls and email attachment scanning because the user manually executes the command. No file is downloaded through inspected channels until after execution begins.
The execution stage is where defenders retain the advantage. ClickFix has a distinctive behavioral fingerprint: a browser-spawned or user-initiated process running an obfuscated, clipboard-sourced command line, frequently via powershell.exe -enc, mshta.exe with remote URLs, or curl | bash patterns on macOS/Linux.
Exploitation Status
Campaigns using these techniques are confirmed active in the wild per Huntress's research. This is not theoretical. There is no CVE and no vendor patch forthcoming — the mitigation is entirely behavioral detection and policy control on the defender side.
Detection & Response
The highest-fidelity detections target the ClickFix execution stage, because that's where the attacker is forced to touch observable telemetry on your endpoint. The Sigma rules below focus on:
- ClickFix-style execution: PowerShell with encoded/obfuscated commands spawned interactively or by script interpreters
mshta.exemaking network connections (a classic ClickFix payload launcher)- Explorer-spawned script interpreters with download cradles (Win+R paste pattern produces explorer.exe as the parent)
---
title: ClickFix-Style Explorer-Spawned PowerShell Download Cradle
id: 3f8a2b71-9c4d-4e6f-a1b2-7d5e9c0f3a84
status: experimental
description: Detects PowerShell spawned by explorer.exe (consistent with Win+R paste execution from ClickFix lures) running download or encoded command patterns. Observed in campaigns abusing AI platform shared content to deliver malware.
references:
- https://www.bleepingcomputer.com/news/security/how-threat-actors-are-turning-trusted-ai-platforms-into-an-attack-surface/
- https://attack.mitre.org/techniques/T1059/001/
- https://attack.mitre.org/techniques/T1204/002/
author: Security Arsenal
date: 2026/01/15
tags:
- attack.execution
- attack.t1059.001
- attack.t1204.002
logsource:
category: process_creation
product: windows
detection:
selection_parent:
ParentImage|endswith: '\explorer.exe'
Image|endswith:
- '\powershell.exe'
- '\pwsh.exe'
- '\powershell_ise.exe'
selection_cmd:
CommandLine|contains:
- ' -enc'
- ' -e '
- 'FromBase64String'
- 'IEX'
- 'Invoke-Expression'
- 'Invoke-WebRequest'
- 'iwr '
- 'curl.exe'
- 'DownloadString'
- 'Start-BitsTransfer'
- 'hidden'
condition: selection_parent and selection_cmd
falsepositives:
- Rare; legitimate admin pasting encoded commands via Run dialog
level: high
---
title: MSHTA Network Connection to External Host
id: 8b1c4d62-5f3a-4e7b-92c1-6a8d0e4f7b35
status: experimental
description: Detects mshta.exe initiating outbound network connections, a common ClickFix payload launcher used in campaigns delivered via weaponized AI platform content and search poisoning.
references:
- https://www.bleepingcomputer.com/news/security/how-threat-actors-are-turning-trusted-ai-platforms-into-an-attack-surface/
- https://attack.mitre.org/techniques/T1218/005/
author: Security Arsenal
date: 2026/01/15
tags:
- attack.defense_evasion
- attack.t1218.005
- attack.command_and_control
logsource:
category: network_connection
product: windows
detection:
selection:
Image|endswith: '\mshta.exe'
filter_rfc1918:
DestinationIp|contains:
- '10.'
- '192.168.'
- '172.16.'
- '172.17.'
- '172.18.'
- '172.19.'
- '172.2'
- '172.30.'
- '172.31.'
- '127.'
condition: selection and not filter_rfc1918
falsepositives:
- Legacy internal HTA applications (should be inventoried and exceptions scoped)
level: high
---
title: Browser-Spawning Script Interpreter After AI Platform Visit
id: 2e7f9a14-3b6c-4d8e-a5f2-9c1b7e0d4a96
status: experimental
description: Detects browser processes spawning cmd, powershell, or mshta — a pattern consistent with ClickFix lures delivered through weaponized Claude Artifacts or malicious shared AI conversations. Browser child processes of this type are abnormal in standard user workflows.
references:
- https://www.bleepingcomputer.com/news/security/how-threat-actors-are-turning-trusted-ai-platforms-into-an-attack-surface/
- https://attack.mitre.org/techniques/T1204/002/
author: Security Arsenal
date: 2026/01/15
tags:
- attack.execution
- attack.t1204.002
logsource:
category: process_creation
product: windows
detection:
selection_parent:
ParentImage|endswith:
- '\chrome.exe'
- '\msedge.exe'
- '\firefox.exe'
- '\brave.exe'
selection_child:
Image|endswith:
- '\cmd.exe'
- '\powershell.exe'
- '\pwsh.exe'
- '\mshta.exe'
- '\wscript.exe'
- '\cscript.exe'
- '\rundll32.exe'
condition: selection_parent and selection_child
falsepositives:
- Browser extensions with native messaging hosts (rare; investigate per-host)
- Enterprise SSO helpers (scoped by known signer)
level: high
For hunting in Microsoft Sentinel / Defender, the query below looks for the Win+R paste signature (explorer-spawned script interpreters with suspicious command lines) and correlates browser telemetry for visits to AI sharing domains in the preceding window:
// Hunt: ClickFix-style execution correlated with AI platform browsing
let Lookback = 7d;
let SuspiciousExec =
DeviceProcessEvents
| where Timestamp > ago(Lookback)
| where InitiatingProcessFileName in~ ("chrome.exe","msedge.exe","firefox.exe","brave.exe")
or InitiatingProcessFileName =~ "explorer.exe"
| where FileName in~ ("powershell.exe","pwsh.exe","mshta.exe","cmd.exe","wscript.exe","cscript.exe","rundll32.exe")
| where ProcessCommandLine has_any ("-enc","FromBase64String","IEX","Invoke-Expression","Invoke-WebRequest","iwr ","DownloadString","Start-BitsTransfer","curl","mshta http")
| project ExecTime=Timestamp, DeviceName, AccountName, FileName, ProcessCommandLine, InitiatingProcessFileName, ReportId;
SuspiciousExec
| join kind=leftouter (
DeviceNetworkEvents
| where Timestamp > ago(Lookback)
| where RemoteUrl has_any ("claude.ai","anthropic.com","chatgpt.com","openai.com","gemini.google.com","copilot.microsoft.com")
| project BrowseTime=Timestamp, DeviceName, RemoteUrl, InitiatingProcessFileName2=InitiatingProcessFileName
) on DeviceName
| where isempty(RemoteUrl) or BrowseTime between (ExecTime .. ExecTime + 30m)
| project ExecTime, DeviceName, AccountName, FileName, ProcessCommandLine, ParentProcess=InitiatingProcessFileName, RemoteUrl, BrowseTime
| order by ExecTime desc
For endpoint forensics during response, Velociraptor can sweep the fleet for the ClickFix execution signature and capture RunMRU evidence — the Run dialog history is a goldmine here because ClickFix victims execute via Win+R, and the pasted command is recorded in the user's RunMRU registry key:
-- Hunt: ClickFix execution artifacts - RunMRU history and suspicious process lineage
-- Pulls Run dialog history (where victims paste ClickFix commands) and
-- cross-references live/recent processes with suspicious parent-child patterns.
LET runmru = SELECT Key.Name AS UserSID,
Data.value AS RunCommand,
Key.Mtime AS LastWrite
FROM read_reg_key(globs='HKEY_USERS/*/Software/Microsoft/Windows/CurrentVersion/Explorer/RunMRU/*')
WHERE RunCommand =~ '(?i)(powershell|mshta|curl|iwr|invoke|base64|http)'
LET procs = SELECT Pid, Ppid, Name, Exe, CommandLine, Username, CreateTime
FROM pslist()
WHERE CommandLine =~ '(?i)(-enc|frombase64string|invoke-expression|downloadstring|mshta http|iwr )'
AND Name =~ '(?i)(powershell|pwsh|mshta|cmd)'
SELECT * FROM runmru
UNION ALL
SELECT UserSID=NULL, RunCommand=format(format='%v | %v', args=[Name, CommandLine]), LastWrite=CreateTime FROM procs
The following PowerShell script helps harden endpoints against ClickFix-style execution and provides a quick audit of RunMRU artifacts on a suspected victim machine:
# ClickFix / AI-Platform-Abuse Hardening & Triage Script
# Run as Administrator. Test in a pilot OU before broad deployment.
# --- 1. Block mshta outbound execution via Windows Defender Attack Surface Reduction ---
# ASR rule: Block execution of potentially obfuscated scripts
Add-MpPreference -AttackSurfaceReductionRules_Ids 5BEB7EFE-FD9A-4556-801D-275E5FFC04CC `
-AttackSurfaceReductionRules_Actions Enabled
# ASR rule: Block process creations from Office/abuse of script interpreters
Add-MpPreference -AttackSurfaceReductionRules_Ids D3E037E1-3EB8-44C8-A917-57927947596D `
-AttackSurfaceReductionRules_Actions Enabled
# --- 2. Restrict PowerShell for standard users via AppLocker policy note ---
# Recommended: deploy WDAC/AppLocker policy constraining powershell.exe to
# admin-only contexts. Flagging current execution policy state for review:
Get-ExecutionPolicy -List | Format-Table Scope, ExecutionPolicy
# --- 3. Triage: dump RunMRU (ClickFix victims execute via Win+R) ---
$runMruPath = 'HKCU:\Software\Microsoft\Windows\CurrentVersion\Explorer\RunMRU'
if (Test-Path $runMruPath) {
Get-ItemProperty $runMruPath | Select-Object * |
Out-File "$env:TEMP\runmru_triage_$(Get-Date -Format yyyyMMdd_HHmmss).txt"
Write-Output "RunMRU exported to $env:TEMP - review for pasted commands referencing http, powershell, mshta, or base64."
}
# --- 4. Audit: recent PowerShell with suspicious command lines in last 24h ---
Get-WinEvent -FilterHashtable @{LogName='Microsoft-Windows-PowerShell/Operational'; StartTime=(Get-Date).AddDays(-1)} -ErrorAction SilentlyContinue |
Where-Object { $_.Message -match 'FromBase64String|Invoke-Expression|DownloadString|mshta|http' } |
Select-Object TimeCreated, Id, Message -First 50 |
Format-List
# --- 5. Verify ASR rule states applied ---
Get-MpPreference | Select-Object -ExpandProperty AttackSurfaceReductionRules_Ids
Remediation and Hardening Guidance
There is no vendor patch — Anthropic cannot "fix" the fact that shared Artifacts exist. The defensive work falls into four buckets:
1. Web filtering — stop trusting domain reputation alone.
- Treat AI sharing/artifact URLs as a distinct URL category in your proxy/SWG. If your filter supports path-level or subresource categorization, alert on (don't necessarily block) AI share-link access from non-developer endpoints.
- Enable TLS inspection where legally permissible so you can inspect Artifact page content for fake error/verification lures, not just the destination domain.
- Block newly registered domains impersonating AI brands; subscribe to a brand-abuse feed for
claude,anthropic,openai,chatgptlookalikes.
2. Search and ad hygiene.
- Block sponsored/ad results at the browser level via enterprise policy where feasible, or train users specifically that sponsored results for software downloads are hostile by default.
- Publish an internal, pinned list of legitimate AI tool download sources. Make the approved path easier than the search-engine path.
3. User awareness — retrain for the ClickFix pattern.
- The single most effective control: no legitimate website, AI assistant, CAPTCHA, or error message will ever ask a user to press Win+R and paste a command. Drill this exact sentence into your awareness program.
- Update phishing simulations to include trusted-domain-hosted lures and ClickFix mechanics — legacy "spot the bad domain" training is now counterproductive for this threat class.
4. Endpoint controls.
- Deploy the ASR rules and Sigma detections above; verify PowerShell Script Block Logging (Event ID 4104) and process creation with command line (Event ID 4688 / Sysmon ID 1) are collected fleet-wide — ClickFix detection is impossible without command-line telemetry.
- Consider disabling or restricting
mshta.exefor standard users via AppLocker/WDAC. - Establish an incident response playbook step: for any suspected ClickFix victim, preserve RunMRU, clipboard history (if enabled), and browser history for the AI platform domains before reimaging.
Governance note: inventory which AI platforms your organization actually sanctions, and route everything else through your existing unsanctioned-SaaS (shadow IT) controls. Attackers are betting that your AI usage policy is vaguer than your other SaaS policies. They're usually right.
Related Resources
Security Arsenal Managed SOC Services AlertMonitor Platform Book a SOC Assessment soc-mdr Intel Hub
Is your security operations ready?
Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.