Three OTX pulses submitted on 2026-11-16 collectively reveal a dangerous and accelerating convergence in the credential theft threat landscape: adversaries are no longer simply harvesting browser cookies and saved passwords — they are systematically targeting AI platform credentials, cloud IAM identities, and LLM session tokens as first-class intelligence commodities.
WardenStealer represents the bleeding edge of this shift. This Rust-based Windows infostealer is purpose-built to exfiltrate authentication tokens, API keys, and local application files from AI platforms including ChatGPT, Claude, Codex, GitHub Copilot, and Cursor. Distributed through ClickFix social engineering lures, cracked software, and malicious advertisements, WardenStealer effectively converts compromised AI sessions into a tradable underground commodity — a technique the research community has labeled LLMjacking. The stolen credentials grant attackers access to enterprise AI pipelines, proprietary codebases connected to Copilot, and sensitive prompts containing intellectual property.
CyberStrikeAI demonstrates what happens after those credentials are acquired. Between September 30 and October 16, 2026, this threat actor executed a high-velocity AWS intrusion spanning four cloud accounts and eight compromised identities. The attack's most alarming characteristic was its machine-speed credential harvesting: one IAM role executed 56 EC2 DescribeInstances API calls in 141 milliseconds, a pace that no human operator could sustain and a clear indicator of automated exploitation tooling. The actor systematically validated stolen credentials, extracted secrets from AWS Secrets Manager, and abused EC2 instance metadata endpoints — a classic cloud kill chain executed at AI-driven velocity.
TypoPrompting completes the picture by revealing the upstream acquisition vector. Threat actors registered 186 typosquatting domains across 27 AI brands — including Claude, ChatGPT, DeepSeek, Perplexity, Midjourney, and ElevenLabs — to intercept users who make simple typing errors when navigating to AI platforms. These domains deliver ClickFix PowerShell lures, Microsoft Store-signed browser-hijacking applications, and search-hijacking Chrome extensions that ultimately deploy infostealers including Aurora Stealer, Net Guard, Invalid Printer, and OctoV2. The campaign targets Finance, Technology, Healthcare, and Telecommunications verticals.
The collective picture is unambiguous: adversaries are building an end-to-end supply chain for AI credential theft — from typosquatting-driven initial access, through infostealer-based harvesting, to machine-speed cloud exploitation using those stolen identities. Enterprise security teams that are not yet treating AI platform credentials and cloud IAM tokens as high-value targets requiring dedicated detection coverage are already behind this curve.
Threat Actor / Malware Profile
WardenStealer
| Attribute | Detail |
|---|---|
| Type | Rust-based Windows Infostealer (MaaS) |
| First Observed | November 2026 |
| Distribution | ClickFix lures, malicious advertisements, cracked software, fake game cheats |
| Primary Targets | ChatGPT, Claude, Codex, GitHub Copilot, Cursor — credentials, auth tokens, cookies, API keys, local app files |
| C2 Infrastructure | wiqld.tattoo, wfave.tattoo (observed) |
| Business Model | Malware-as-a-Service — enables LLMjacking and AI session hijacking at scale |
WardenStealer's Rust-based architecture provides several operational advantages for its operators: Rust binaries are notoriously difficult for traditional AV engines to analyze due to complex inlining, aggressive compiler optimization, and a lack of mature reverse engineering tooling compared to C/C++ binaries. The malware follows the standard infostealer playbook — initial execution via ClickFix social engineering (which tricks users into pasting malicious PowerShell commands into Run dialogs or terminals), followed by local data harvesting from browser credential stores, application data directories, and AI client configuration files.
The malware's specific targeting of AI application local storage represents a meaningful evolution. Tools like Cursor and GitHub Copilot store session tokens, API keys, and cached context data locally. Exfiltrating these artifacts provides attackers with not just credentials, but potential access to codebases, private repositories, and enterprise AI usage patterns.
CyberStrikeAI
| Attribute | Detail |
|---|---|
| Type | Automated Cloud Intrusion Platform |
| Campaign Window | 2026-09-30 to 2026-10-16 |
| Scope | 4 AWS accounts, 8 compromised identities |
| Key Technique | Machine-speed credential harvesting — 56 EC2 API calls in 141ms |
| Cloud Services Abused | EC2 (DescribeInstances, User Data), Secrets Manager, IAM, SSM, Instance Metadata Service (IMDS) |
| Observed C2/Origin IPs | 43.240.29.33, 43.240.29.28, 43.240.29.11 |
CyberStrikeAI operates as an automated intrusion framework — likely AI-assisted given the speed and breadth of its operations. The attack chain follows a deliberate pattern: validate compromised credentials via low-risk API calls (DescribeInstances, GetCallerIdentity), enumerate the environment's EC2 instances and extract user data scripts (which frequently contain hardcoded credentials and bootstrap secrets), pivot to Secrets Manager to harvest stored credentials, and then leverage SSM for lateral movement across managed instances.
The use of IMDS exploitation is particularly noteworthy. By querying the EC2 Instance Metadata Service from compromised instances, the attacker can harvest temporary IAM role credentials — a technique that bypasses traditional credential storage controls entirely. The machine-speed API call pattern (56 calls in 141ms) strongly suggests scripted automation with pre-built playbooks rather than interactive operator activity.
TypoPrompting Campaign (Multiple Malware Families)
| Attribute | Detail |
|---|---|
| Type | Typosquatting-driven Malware Distribution Campaign |
| Scale | 186 domains across 27 AI brands |
| Brands Impersonated | Claude, ChatGPT, DeepSeek, Perplexity, Midjourney, ElevenLabs, Runway ML, and others |
| Payloads | Aurora Stealer, Net Guard, Invalid Printer, OctoV2 |
| Delivery Methods | ClickFix PowerShell lures, Microsoft Store-signed browser hijackers, search-hijacking Chrome extensions |
| Targeted Industries | Finance, Technology, Healthcare, Telecommunications |
The TypoPrompting campaign exploits a deceptively simple human behavior: typing errors when navigating to AI platform URLs. Domains like opnai.com (missing the 'e' in OpenAI), chtgpt.com (missing the 'a' in ChatGPT), anthropiv.com (substituting 'v' for 'c' in Anthropic), cloude.ai (substituting 'ou' for 'au' in Claude), midj0urney.org (zero for 'o' in Midjourney), and chatlgpt.com (inserted 'l' in ChatGPT) are crafted to intercept these typos.
Once a user lands on a typosquatting domain, the infection chain branches into multiple delivery mechanisms. The ClickFix approach presents fake CAPTCHA or error pages instructing users to press Windows+R and paste a clipboard-injected PowerShell command — this is the same social engineering vector used to deliver WardenStealer. Alternatively, users may be prompted to install what appears to be a legitimate browser extension or Microsoft Store application, which then hijacks browser search settings, injects advertisements, or directly harvests credentials.
Aurora Stealer is a well-documented Go-based infostealer that targets browser credentials, cryptocurrency wallets, and messaging application tokens. Net Guard operates as a browser-hijacking application, often distributed through the Microsoft Store with valid code signing to evade reputation-based detection. OctoV2 is an Android malware family that has been adapted for cross-platform delivery. Invalid Printer leverages social engineering around printer driver installation prompts.
IOC Analysis
Indicator Breakdown
| Type | Count | Source Pulse | Operationalization |
|---|---|---|---|
| FileHash-SHA256 | 75 | WardenStealer | Block in EDR/AV. Import into threat intel platform. Hash-based detection is point-in-time — Rust binaries are easily recompiled with different hashes. |
| Domain | 10 | WardenStealer (2), TypoPrompting (8) | DNS sinkhole/block. The .tattoo TLD for WardenStealer C2 is unusual and should be monitored. Typosquatting domains are static and highly blockable. |
| IPv4 | 3 | CyberStrikeAI | Block at perimeter. Monitor CloudTrail for API calls originating from these IPs. Add to AWS WAF and GuardDuty threat lists. |
SOC Operationalization Guidance
File Hashes (SHA256): The 75 WardenStealer hashes should be ingested into your EDR platform's block list immediately. However, hash-based detection alone is insufficient for Rust-based malware — the compiler's optimization pipeline means that even minor source code changes produce entirely different binaries. Use these hashes as an initial sweep and retro-hunt mechanism, but invest primarily in behavioral detection.
Domains: The typosquatting domains are highly actionable — they are static, brand-specific, and unlikely to have legitimate business use. Implement DNS-layer blocking for all listed domains. Beyond the specific IOCs, implement a permutation detection feed (e.g., via DNSTwist or similar tooling) that continuously monitors for newly registered domains that are one-edit-distance from your organization's critical AI platform domains.
The WardenStealer C2 domains (wiqld.tattoo, wfave.tattoo) use the .tattoo TLD, which is uncommon in enterprise environments. Alert on any DNS resolution to .tattoo TLD domains as a hunting heuristic, while acknowledging this is a broad signal that requires triage.
IPv4 Addresses: The three CyberStrikeAI IPs (43.240.29.33, 43.240.29.28, 43.240.29.11) belong to the 43.240.29.0/24 range. Consider blocking the full /24 at the perimeter after validating no legitimate business traffic originates from this range. More importantly, retro-hunt CloudTrail logs for any API calls originating from these IPs across all AWS accounts — if found, those API calls represent confirmed unauthorized access.
Tooling: Use MISP or your TIP for IOC management and correlation. Run hashes through VirusTotal/Hybrid Analysis for behavioral reports. Use DNSTwist for typosquatting permutation monitoring. Leverage AWS GuardDuty's threat intelligence feeds alongside custom threat lists containing the CyberStrikeAI IPs.
Detection Engineering
---
title: WardenStealer AI Credential Theft — ClickFix Execution and AI Platform Data Access
id: 8f3a2b1c-4d5e-6f7a-8b9c-0d1e2f3a4b5c
status: experimental
description: Detects WardenStealer infostealer execution patterns including ClickFix PowerShell social engineering, AI application credential access, and C2 communication to known WardenStealer infrastructure.
author: Security Arsenal Threat Intelligence
references:
- https://cyberint.com/blog/dark-web/wardenstealer-malware-turning-ai-access-into-an-underground-commodity/
date: 2026/11/16
tags:
- attack.credential_access
- attack.t1552
- attack.t1555
- attack.t1059.001
- attack.t1071.001
logsource:
category: process_creation
product: windows
detection:
selection_clickfix_parent:
ParentImage|endswith:
- '\explorer.exe'
- '\msedge.exe'
- '\chrome.exe'
- '\firefox.exe'
- '\brave.exe'
selection_clickfix_cmd:
CommandLine|contains:
- 'powershell'
- 'pwsh'
selection_clickfix_flags:
CommandLine|contains:
- '-enc'
- '-encodedcommand'
- '-w hidden'
- '-windowstyle hidden'
- '-ep bypass'
- '-executionpolicy bypass'
- 'IEX'
- 'Invoke-Expression'
- 'DownloadString'
- 'Net.WebClient'
selection_ai_target:
CommandLine|contains:
- '\OpenAI\'
- '\ChatGPT\'
- '\Claude\'
- '\Copilot\'
- '\Cursor\'
- '\Codex\'
- 'copilot'
- 'openai'
- 'anthropic'
- '.cursor'
- 'claude_desktop'
selection_c2_domain:
CommandLine|contains:
- 'wiqld.tattoo'
- 'wfave.tattoo'
condition: >
(selection_clickfix_parent and selection_clickfix_cmd and selection_clickfix_flags)
or selection_ai_target
or selection_c2_domain
falsepositives:
- Legitimate IT administration using PowerShell with bypass flags
- Developer tooling accessing AI platform configuration files
level: high
---
title: CyberStrikeAI AWS Cloud Credential Harvesting — Machine-Speed API Enumeration
id: 1a2b3c4d-5e6f-7a8b-9c0d-1e2f3a4b5c6d
status: experimental
description: Detects anomalous AWS API call velocity consistent with CyberStrikeAI automated credential harvesting. Identifies rapid EC2 DescribeInstances calls, Secrets Manager access, and IMDS credential theft patterns from suspicious source IPs.
author: Security Arsenal Threat Intelligence
references:
- https://www.elastic.co/security-labs/threat-command/cyberstrikeai-aws-cloudtrail-intrusion
date: 2026/11/16
tags:
- attack.credential_access
- attack.t1552.005
- attack.t1528
- attack.discovery
- attack.t1580
logsource:
product: aws
service: cloudtrail
detection:
selection_ec2_enum:
eventSource: 'ec2.amazonaws.com'
eventName:
- 'DescribeInstances'
- 'DescribeInstanceAttribute'
- 'GetConsoleOutput'
- 'DescribeSnapshots'
selection_secrets:
eventSource: 'secretsmanager.amazonaws.com'
eventName:
- 'GetSecretValue'
- 'ListSecrets'
- 'DescribeSecret'
- 'BatchGetSecretValue'
selection_iam_enum:
eventSource: 'iam.amazonaws.com'
eventName:
- 'GetCallerIdentity'
- 'ListAccessKeys'
- 'ListRoles'
- 'ListUsers'
- 'GetAccessKeyLastUsed'
selection_ssm:
eventSource: 'ssm.amazonaws.com'
eventName:
- 'SendCommand'
- 'StartSession'
- 'GetParameter'
- 'GetParameters'
selection_suspicious_ip:
sourceIPAddress:
- '43.240.29.33'
- '43.240.29.28'
- '43.240.29.11'
timeframe: 5m
condition: >
selection_ec2_enum or selection_secrets or selection_iam_enum or selection_ssm or selection_suspicious_ip
falsepositives:
- Automated infrastructure-as-code pipelines (Terraform, Pulumi)
- Cloud security posture management (CSPM) tools
- Legitimate DevOps automation
level: critical
---
title: TypoPrompting AI Brand Typosquatting — Malicious Domain Resolution and Payload Delivery
id: 7c8d9e0f-1a2b-3c4d-5e6f-7a8b9c0d1e2f
status: experimental
description: Detects DNS resolution and network connections to known TypoPrompting typosquatting domains impersonating AI platforms including ChatGPT, Claude, Midjourney, and others. May indicate ClickFix lure delivery or infostealer payload download.
author: Security Arsenal Threat Intelligence
references:
- https://www.island.io/blog/typoprompting-from-ai-typos-to-malware
date: 2026/11/16
tags:
- attack.initial_access
- attack.t1189
- attack.t1566
- attack.command_and_control
- attack.t1071
logsource:
category: dns
product: windows
detection:
selection_typosquat:
query|contains:
- 'opnai.com'
- 'runawayml.com'
- 'griok.com'
- 'anthropiv.com'
- 'midj0urney.org'
- 'chtgpt.com'
- 'cloude.ai'
- 'chatlgpt.com'
- 'wiqld.tattoo'
- 'wfave.tattoo'
condition: selection_typosquat
falsepositives:
- Security research or threat hunting activity
- DNS filtering validation testing
level: high
// ============================================================================
// Security Arsenal — AI Credential Theft & Cloud Intrusion Hunt Query
// Covers: WardenStealer, CyberStrikeAI, TypoPrompting campaign indicators
// Platform: Microsoft Sentinel / Microsoft 365 Defender
// Timeframe: Last 30 days (adjust as needed)
// ============================================================================
let typoPromptingDomains = dynamic([
"opnai.com", "runawayml.com", "griok.com", "anthropiv.com",
"midj0urney.org", "chtgpt.com", "cloude.ai", "chatlgpt.com"
]);
let wardenStealerC2 = dynamic([
"wiqld.tattoo", "wfave.tattoo"
]);
let cyberStrikeAIIPs = dynamic([
"43.240.29.33", "43.240.29.28", "43.240.29.11"
]);
let wardenStealerHashes = dynamic([
"2174253d9e78ed616c82cf2c7334817b3af400cf98c9ac2f348aedb0801d280a",
"b7f4cbde98525a79c38d64322dba7d9d2289e4903ad5a245e80ea3a6cfcc7ef5",
"b797e3534789157cff3ffb1299905648315d89316ff355d122c57e475be1a240",
"d9bb4cb6977d817979c51558d4e12da10b747ed87122b219a62a1ce64576da92",
"ffbb2d190c87f9832e705692d0f9bba7a885672b842a8503279ac50bf0eb0a69",
"4b1178a7cf24888437012e347ba87329d9caa0caee75d9543b7e6e0ed03a7e70"
]);
// --- Hunt 1: DNS/Network connections to typosquatting and C2 domains ---
let DomainHits = DeviceNetworkEvents
| where TimeGenerated > ago(30d)
| where RemoteUrl has_any (typoPromptingDomains) or RemoteUrl has_any (wardenStealerC2)
| project TimeGenerated, DeviceName, RemoteUrl, RemoteIP, InitiatingProcessFileName,
InitiatingProcessCommandLine, ActionType, RemotePort
| extend HuntSource = "DomainResolution";
// --- Hunt 2: Network connections to CyberStrikeAI IPs ---
let IPHits = DeviceNetworkEvents
| where TimeGenerated > ago(30d)
| where RemoteIP in (cyberStrikeAIIPs)
| project TimeGenerated, DeviceName, RemoteUrl, RemoteIP, InitiatingProcessFileName,
InitiatingProcessCommandLine, ActionType, RemotePort
| extend HuntSource = "CyberStrikeAI_IP";
// --- Hunt 3: File hash matches for WardenStealer ---
let HashHits = DeviceFileEvents
| where TimeGenerated > ago(30d)
| where SHA256 in (wardenStealerHashes)
| project TimeGenerated, DeviceName, FileName, FolderPath, SHA256,
InitiatingProcessFileName, InitiatingProcessCommandLine, ActionType
| extend HuntSource = "WardenStealer_Hash";
// --- Hunt 4: ClickFix-style PowerShell execution (browser-spawned encoded commands) ---
let ClickFixHits = DeviceProcessEvents
| where TimeGenerated > ago(30d)
| where FileName in~ ("powershell.exe", "pwsh.exe")
| where InitiatingProcessFileName in~ (
"explorer.exe", "msedge.exe", "chrome.exe", "firefox.exe", "brave.exe", "mshta.exe", "rundll32.exe"
)
| where ProcessCommandLine has_any (
"-enc", "-encodedcommand", "-w hidden", "-windowstyle hidden",
"-ep bypass", "-executionpolicy bypass", "IEX", "Invoke-Expression",
"DownloadString", "Net.WebClient", "Start-BitsTransfer"
)
| project TimeGenerated, DeviceName, FileName, ProcessCommandLine,
InitiatingProcessFileName, InitiatingProcessCommandLine, AccountName
| extend HuntSource = "ClickFix_PowerShell";
// --- Hunt 5: Access to AI application credential/config directories ---
let AICredentialAccess = DeviceFileEvents
| where TimeGenerated > ago(30d)
| where FolderPath has_any (
"\\OpenAI\\", "\\ChatGPT\\", "\\Claude\\", "\\Copilot\\",
"\\Cursor\\", "\\.cursor\\", "\\claude"
)
| where InitiatingProcessFileName !in~ (
"msedge.exe", "chrome.exe", "Code.exe", "cursor.exe",
"explorer.exe", "SearchIndexer.exe", "MsMpEng.exe", "claude.exe"
)
| project TimeGenerated, DeviceName, FileName, FolderPath,
InitiatingProcessFileName, InitiatingProcessCommandLine, ActionType
| extend HuntSource = "AI_Credential_Access";
// --- Combine all hunts ---
union isfuzzy=true DomainHits, IPHits, HashHits, ClickFixHits, AICredentialAccess
| sort by TimeGenerated desc
| extend AlertPriority = case(
HuntSource == "CyberStrikeAI_IP", "CRITICAL",
HuntSource == "WardenStealer_Hash", "HIGH",
HuntSource == "DomainResolution", "HIGH",
HuntSource == "ClickFix_PowerShell", "MEDIUM",
HuntSource == "AI_Credential_Access", "MEDIUM",
"LOW"
)
# ============================================================================
# Security Arsenal — AI Credential Theft IOC Hunt Script
# Targets: WardenStealer, TypoPrompting, CyberStrikeAI artifacts
# Platform: Windows 10/11 endpoints
# Run as: Administrator (for registry and system-level checks)
# ============================================================================
$ErrorActionPreference = 'SilentlyContinue'
$report = @()
$timestamp = Get-Date -Format "yyyy-MM-dd_HH-mm-ss"
Write-Host "[+] Security Arsenal AI Credential Theft Hunt — $timestamp" -ForegroundColor Cyan
Write-Host "[=] Checking WardenStealer file hash indicators..." -ForegroundColor Yellow
# --- Check 1: WardenStealer SHA256 hashes in common staging directories ---
$malHashes = @(
"2174253d9e78ed616c82cf2c7334817b3af400cf98c9ac2f348aedb0801d280a",
"b7f4cbde98525a79c38d64322dba7d9d2289e4903ad5a245e80ea3a6cfcc7ef5",
"b797e3534789157cff3ffb1299905648315d89316ff355d122c57e475be1a240",
"d9bb4cb6977d817979c51558d4e12da10b747ed87122b219a62a1ce64576da92",
"ffbb2d190c87f9832e705692d0f9bba7a885672b842a8503279ac50bf0eb0a69",
"4b1178a7cf24888437012e347ba87329d9caa0caee75d9543b7e6e0ed03a7e70"
)
$searchPaths = @(
"$env:TEMP",
"$env:APPDATA",
"$env:LOCALAPPDATA",
"$env:USERPROFILE\Downloads",
"$env:USERPROFILE\Desktop",
"$env:ProgramData"
)
foreach ($path in $searchPaths) {
if (Test-Path $path) {
$files = Get-ChildItem -Path $path -Recurse -File -Depth 3 -ErrorAction SilentlyContinue |
Where-Object { $_.Extension -in '.exe','.dll','.ps1','.bat','.msi','.scr','.com' } |
Select-Object -First 500
foreach ($file in $files) {
$hash = (Get-FileHash -Path $file.FullName -Algorithm SHA256 -ErrorAction SilentlyContinue).Hash
if ($hash -and ($malHashes -contains $hash.ToLower())) {
$report += [PSCustomObject]@{
Check = "WardenStealer_Hash"
Status = "FOUND"
Path = $file.FullName
Hash = $hash
Detail = "CRITICAL: Known WardenStealer binary detected"
}
Write-Host " [!] FOUND: $($file.FullName) — Hash: $hash" -ForegroundColor Red
}
}
}
}
Write-Host " [✓] Hash scan complete." -ForegroundColor Green
# --- Check 2: DNS cache for typosquatting and C2 domains ---
Write-Host "[=] Checking DNS cache for typosquatting and C2 domains..." -ForegroundColor Yellow
$malDomains = @(
"opnai.com", "runawayml.com", "griok.com", "anthropiv.com",
"midj0urney.org", "chtgpt.com", "cloude.ai", "chatlgpt.com",
"wiqld.tattoo", "wfave.tattoo"
)
$dnsCache = Get-DnsClientCache -ErrorAction SilentlyContinue
foreach ($domain in $malDomains) {
$match = $dnsCache | Where-Object { $_.Entry -like "*$domain*" }
if ($match) {
$report += [PSCustomObject]@{
Check = "DNS_Cache_Match"
Status = "FOUND"
Path = "DNS Cache"
Hash = "N/A"
Detail = "HIGH: Domain $domain found in DNS cache — possible typosquat visit"
}
Write-Host " [!] FOUND: $domain in DNS cache" -ForegroundColor Red
}
}
Write-Host " [✓] DNS cache check complete." -ForegroundColor Green
# --- Check 3: Network connections to CyberStrikeAI IPs ---
Write-Host "[=] Checking active/historical network connections to CyberStrikeAI IPs..." -ForegroundColor Yellow
$csIPs = @("43.240.29.33", "43.240.29.28", "43.240.29.11")
$connections = Get-NetTCPConnection -ErrorAction SilentlyContinue
foreach ($ip in $csIPs) {
$match = $connections | Where-Object { $_.RemoteAddress -eq $ip }
if ($match) {
foreach ($conn in $match) {
$proc = Get-Process -Id $conn.OwningProcess -ErrorAction SilentlyContinue
$report += [PSCustomObject]@{
Check = "CyberStrikeAI_IP"
Status = "FOUND"
Path = "$($proc.Path)"
Hash = "N/A"
Detail = "CRITICAL: Active connection to $ip from $($proc.ProcessName) (PID: $($conn.OwningProcess))"
}
Write-Host " [!] FOUND: Connection to $ip from $($proc.ProcessName)" -ForegroundColor Red
}
}
}
Write-Host " [✓] Network connection check complete." -ForegroundColor Green
# --- Check 4: Suspicious Run key persistence entries ---
Write-Host "[=] Checking registry Run keys for suspicious persistence..." -ForegroundColor Yellow
$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",
"HKLM:\SOFTWARE\WOW6432Node\Microsoft\Windows\CurrentVersion\Run"
)
$suspiciousKeywords = @("warden", "stealer", "update", "service", "helper", "sync", "agent")
foreach ($key in $runKeys) {
if (Test-Path $key) {
$values = Get-ItemProperty -Path $key -ErrorAction SilentlyContinue
$values.PSObject.Properties | ForEach-Object {
$name = $_.Name
$value = $_.Value
if ($name -notmatch "^(PSPath|PSParentPath|PSChildName|PSDrive|PSProvider)$") {
# Flag entries pointing to TEMP, AppData, or Public directories
if ($value -match "(\\Temp\\|\\AppData\\|\\Public\\|\\ProgramData\\.*\.exe)") {
$report += [PSCustomObject]@{
Check = "Registry_Persistence"
Status = "SUSPICIOUS"
Path = "$key\$name"
Hash = "N/A"
Detail = "MEDIUM: Run key entry '$name' points to non-standard location: $value"
}
Write-Host " [?] SUSPICIOUS: $key\$name = $value" -ForegroundColor Magenta
}
}
}
}
}
Write-Host " [✓] Registry persistence check complete." -ForegroundColor Green
# --- Check 5: Scheduled tasks with suspicious characteristics ---
Write-Host "[=] Checking scheduled tasks for suspicious entries..." -ForegroundColor Yellow
$tasks = Get-ScheduledTask -ErrorAction SilentlyContinue | Where-Object {
$_.TaskPath -notlike "\Microsoft*" -and $_.State -ne "Disabled"
}
foreach ($task in $tasks) {
$actions = $task.Actions
foreach ($action in $actions) {
if ($action.Execute -match "(powershell|cmd|mshta|rundll32|wscript|cscript)" -and
$action.Arguments -match "(-enc|-w hidden|bypass|IEX|DownloadString|hidden)") {
$report += [PSCustomObject]@{
Check = "Scheduled_Task"
Status = "SUSPICIOUS"
Path = $task.TaskName
Hash = "N/A"
Detail = "HIGH: Scheduled task '$($task.TaskName)' executes $($action.Execute) with suspicious args: $($action.Arguments)"
}
Write-Host " [?] SUSPICIOUS: Task '$($task.TaskName)' — $($action.Execute) $($action.Arguments)" -ForegroundColor Magenta
}
}
}
Write-Host " [✓] Scheduled task check complete." -ForegroundColor Green
# --- Check 6: Browser extension artifacts (TypoPrompting search hijackers) ---
Write-Host "[=] Checking Chrome extension directories for suspicious entries..." -ForegroundColor Yellow
$chromeExtPath = "$env:LOCALAPPDATA\Google\Chrome\User Data\Default\Extensions"
$edgeExtPath = "$env:LOCALAPPDATA\Microsoft\Edge\User Data\Default\Extensions"
$knownBadExtensions = @("net guard", "invalid printer", "search helper", "web protect")
foreach ($extBase in @($chromeExtPath, $edgeExtPath)) {
if (Test-Path $extBase) {
$extensions = Get-ChildItem -Path $extBase -Directory -ErrorAction SilentlyContinue
foreach ($ext in $extensions) {
$manifestPath = Get-ChildItem -Path $ext.FullName -Recurse -Filter "manifest.json" -Depth 2 -ErrorAction SilentlyContinue | Select-Object -First 1
if ($manifestPath) {
$manifest = Get-Content $manifestPath.FullName -Raw | ConvertFrom-Json -ErrorAction SilentlyContinue
$extName = $manifest.name
foreach ($bad in $knownBadExtensions) {
if ($extName -match $bad) {
$report += [PSCustomObject]@{
Check = "Browser_Extension"
Status = "FOUND"
Path = $ext.FullName
Hash = "N/A"
Detail = "HIGH: Suspicious extension '$extName' matching '$bad' pattern"
}
Write-Host " [!] FOUND: Extension '$extName' at $($ext.FullName)" -ForegroundColor Red
}
}
# Flag extensions requesting broad permissions
if ($manifest.permissions -match "(cookies|webRequest|tabs|<all_urls>|management)") {
$report += [PSCustomObject]@{
Check = "Browser_Extension_Perms"
Status = "SUSPICIOUS"
Path = $ext.FullName
Hash = "N/A"
Detail = "MEDIUM: Extension '$extName' has broad permissions: $($manifest.permissions -join ', ')"
}
}
}
}
}
}
Write-Host " [✓] Browser extension check complete." -ForegroundColor Green
# --- Check 7: AI application credential file access artifacts ---
Write-Host "[=] Checking for AI application credential/config directories..." -ForegroundColor Yellow
$aiAppPaths = @(
"$env:APPDATA\OpenAI",
"$env:APPDATA\ChatGPT",
"$env:APPDATA\Claude",
"$env:APPDATA\Cursor",
"$env:USERPROFILE\.cursor",
"$env:APPDATA\GitHub Copilot",
"$env:LOCALAPPDATA\claude-cli"
)
foreach ($aiPath in $aiAppPaths) {
if (Test-Path $aiPath) {
Write-Host " [i] AI app directory exists: $aiPath" -ForegroundColor Blue
# Check for recently modified files (potential tampering/exfil indicator)
$recentFiles = Get-ChildItem -Path $aiPath -Recurse -File -ErrorAction SilentlyContinue |
Where-Object { $_.LastWriteTime -gt (Get-Date).AddDays(-7) }
if ($recentFiles) {
foreach ($rf in $recentFiles) {
Write-Host " [?] Recently modified: $($rf.FullName) — $($rf.LastWriteTime)" -ForegroundColor Magenta
}
}
}
}
Write-Host " [✓] AI application directory check complete." -ForegroundColor Green
# --- Summary ---
Write-Host "`n========================================" -ForegroundColor Cyan
Write-Host " HUNT COMPLETE — Results Summary" -ForegroundColor Cyan
Write-Host "========================================" -ForegroundColor Cyan
$foundCount = ($report | Where-Object { $_.Status -eq "FOUND" }).Count
$suspCount = ($report | Where-Object { $_.Status -eq "SUSPICIOUS" }).Count
Write-Host " CRITICAL/FOUND: $foundCount" -ForegroundColor $(if($foundCount -gt 0){"Red"}else{"Green"})
Write-Host " SUSPICIOUS: $suspCount" -ForegroundColor $(if($suspCount -gt 0){"Magenta"}else{"Green"})
if ($report.Count -gt 0) {
$outFile = "$env:USERPROFILE\Desktop\AI_Credential_Theft_Hunt_$timestamp.csv"
$report | Export-Csv -Path $outFile -NoTypeInformation
Write-Host "`n Report saved to: $outFile" -ForegroundColor Yellow
} else {
Write-Host " No indicators found. System appears clean for checked IOCs." -ForegroundColor Green
}
Write-Host "`n[+] Hunt complete. Review any FOUND or SUSPICIOUS items immediately." -ForegroundColor Cyan
Response Priorities
Immediate (0–4 Hours)
-
Block all IOCs at the network perimeter. Push the eight typosquatting domains, two WardenStealer C2 domains, and three CyberStrikeAI IPs to your DNS filtering layer, web proxy, and firewall block lists. Consider blocking the full
43.240.29.0/24range after validating no legitimate traffic. -
Push SHA256 hashes to EDR block lists. All 75 WardenStealer hashes should be ingested into your endpoint protection platform's custom block list. Trigger a retroactive hash sweep across all endpoints.
-
Deploy Sigma and KQL detections. The Sigma rules and KQL queries provided above should be deployed to your SIEM and EDR platforms immediately. Tune false positive thresholds based on your environment's baseline.
-
Run the PowerShell hunt script on high-priority endpoints — particularly developer workstations, AI/ML engineering machines, and systems with access to cloud consoles.
-
CloudTrail retro-hunt. Query AWS CloudTrail across all accounts for API calls originating from
43.240.29.33,43.240.29.28, and43.240.29.11over the past 90 days. Any hits represent confirmed unauthorized access requiring immediate incident response activation.
24 Hours
-
Force credential rotation for all AI platform accounts. If any endpoint shows evidence of WardenStealer or TypoPrompting compromise, immediately rotate API keys, session tokens, and passwords for ChatGPT, Claude, GitHub Copilot, Cursor, and any other AI platform in use. Do not rely on users self-reporting — assume exposure.
-
Rotate AWS credentials. If CyberStrikeAI IPs appear in CloudTrail, rotate all IAM access keys, invalidate all active sessions, review Secrets Manager for unauthorized GetSecretValue calls, and audit EC2 user data scripts for exposed credentials.
-
Browser extension audit. Enumerate all installed browser extensions across the fleet. Remove any extensions matching the TypoPrompting campaign's known payloads (Net Guard, Invalid Printer). Implement an extension allowlist policy.
-
Identity verification. For any user whose endpoint shows infostealer indicators, require re-enrollment in MFA, review recent authentication logs for anomalous sessions, and check for unauthorized OAuth grants in Microsoft 365 and Google Workspace.
-
Assess AI data exposure. If AI platform credentials were compromised, evaluate what data may have been accessible through those sessions. Codebases connected to Copilot, proprietary prompts, and AI-assisted conversations may contain sensitive intellectual property.
One Week
-
Implement AI credential governance. Establish a formal policy for AI platform credential management: enterprise accounts only (no personal accounts on corporate devices), API key rotation schedules, and monitoring for unauthorized AI platform access.
-
Deploy DNS-layer typosquatting detection. Implement a DNSTwist-based monitoring pipeline or commercial brand protection service to alert on newly registered domains that are one-edit-distance from AI platforms your organization uses. Block these proactively.
-
Harden AWS against machine-speed attacks. Implement SCPs (Service Control Policies) to restrict Secrets Manager access to specific roles. Enable IMDSv2 enforcement across all EC2 instances to prevent metadata credential theft. Configure CloudTrail anomaly detection alerts for API call velocity thresholds (e.g., alert if any single identity makes more than 10 EC2 DescribeInstances calls per second).
-
Anti-ClickFix controls. Implement browser-level protections against clipboard injection. Consider Group Policy restrictions on Windows+R for non-admin users, or deploy endpoint rules that block PowerShell spawned from browsers with encoded or obfuscated arguments.
-
User awareness training. Brief all staff — especially developers and AI platform users — on ClickFix social engineering (never paste commands from web pages into Run dialogs), typosquatting risks (bookmark AI platform URLs rather than typing them), and the value of AI credentials to adversaries.
-
Threat hunting cadence. Incorporate the KQL queries and Sigma rules from this briefing into your standing threat hunting playbook. Run the AI credential access hunt weekly. Monitor for new OTX pulses related to WardenStealer, CyberStrikeAI, and TypoPrompting.
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.