Skip to navigation

Threat reportSupply ChainTL-2026-1018

Phantom Squatting: Adversaries Weaponize AI-Hallucinated Domains as Supply Chain Attack Vector

highACTIVE

Unit 42 (Palo Alto Networks) documents 'phantom squatting,' where adversaries systematically probe LLMs to discover domains the models hallucinate for legitimate brands, then preemptively register those domains before defenders detect them. Analysis of 913 global brands across 685,339 adversarial prompts against two production LLMs generated 2.1 million unique URLs, of which 13,229 (0.61%) were confirmed malicious and roughly 250,000 unregistered phantom domains remain immediate registration opportunities for attackers.

Key facts

IdentifiersTL-2026-1018
Also known asPhantom Squatting, AI Hallucination Squatting, LLM Domain Hallucination Squatting
SeverityHIGH
Affected productsMultiple (913 global brands) Brand-facing customer portals, e-commerce marketplaces, postal/logistics services, banking portals, sports-betting platforms; Unnamed LLM vendors LLM1 (production-optimized mini-class enterprise model, April 2025 build); LLM2 (low-latency lite-class frontier model, June 2025 build)
MalwareMontana Empire
ToolsTelegram-based C2 relay panel (admin panel branded 'Kimseye Güvenme' / 'Trust No One')
MITRE ATT&CK techniquesT1005 Data from Local System, T1040 Network Sniffing, T1071 Application Layer Protocol, T1102 Web Service, T1187 Forced Authentication, T1199 Trusted Relationship, T1204 User Execution, T1566 Phishing, T1583 Acquire Infrastructure, T1587 Develop Capabilities, T1589 Gather Victim Identity Information, T1590 Gather Victim Network Information and 4 more in the attack flow
MotivationFINANCIAL
Target sectorstechnology, finance, health, ecommerce, government administration, gambling, logistics, postal, banking
Target regionsGlobal, South Asia, Europe, Middle East
Detection rules9 (1 shown free)
Sources7 cited sources

Detection coverage

As of 2026-06-30, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-1018 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: Autonomous Agent/Scripting Engine Fetches Newly-Registered Zero-Reputation Domain

Detects PowerShell/pwsh (or other automation runtimes) issuing Invoke-WebRequest/Invoke-RestMethod calls to a domain with no prior allowlist/reputation history — the phantom-squatting exploitation pattern where an AI agent trusts an LLM-hallucinated URL without human verification (e.g. the letgovip.zip / Montana Empire archive pull).

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1204, T1199, T1583, T1590
  • Data sources: sysmon, windows_security
  • Severity: high
  • Confidence: high

Splunk SPL

index=sysmon sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" EventCode=1
| where match(Image, "(?i)\\\\(powershell|pwsh)\.exe$")
| where match(CommandLine, "(?i)invoke-(webrequest|restmethod)")
| rex field=CommandLine "(?i)(https?:\/\/(?<fetch_domain>[a-zA-Z0-9\.\-]+))"
| where isnotnull(fetch_domain)
| lookup allowlisted_domains.csv domain AS fetch_domain OUTPUT is_allowlisted
| where isnull(is_allowlisted)
| lookup domain_first_seen.csv domain AS fetch_domain OUTPUT first_seen_days_ago
| where first_seen_days_ago < 30 OR isnull(first_seen_days_ago)
| stats count min(_time) as first_seen max(_time) as last_seen values(CommandLine) as commands by Computer, User, fetch_domain
| where count > 0
| sort -last_seen

Microsoft KQL

DeviceProcessEvents
| where Timestamp > ago(24h)
| where FileName in~ ("powershell.exe", "pwsh.exe")
| where ProcessCommandLine has_any ("Invoke-WebRequest", "Invoke-RestMethod")
| extend FetchDomain = extract(@"https?://([a-zA-Z0-9\.\-]+)", 1, ProcessCommandLine)
| where isnotempty(FetchDomain)
| join kind=leftanti (AllowlistedDomains) on $left.FetchDomain == $right.Domain
| project Timestamp, DeviceName, AccountName, FileName, ProcessCommandLine, FetchDomain
| summarize FirstSeen=min(Timestamp), LastSeen=max(Timestamp), Count=count() by DeviceName, AccountName, FetchDomain
| order by LastSeen desc

Sigma

title: Unverified Outbound Fetch to Newly-Observed Domain by Automation/Agent Process
id: 7c1e2f3a-4b5c-4d6e-8f9a-1b2c3d4e5f60
status: experimental
description: Detects PowerShell/pwsh issuing a web fetch to a domain not present in the organizational allowlist, consistent with an autonomous AI agent acting on an LLM-hallucinated (phantom-squatted) URL without human-in-the-loop verification.
references:
  - https://unit42.paloaltonetworks.com/phantom-squatting-hallucinated-web-domains/
author: Threadlinqs Team
date: 2026-06-30
tags:
  - attack.reconnaissance
  - attack.t1590
  - attack.initial-access
  - attack.t1199
  - attack.execution
  - attack.t1204
  - attack.resource-development
  - attack.t1583
logsource:
  product: windows
  category: process_creation
detection:
  selection:
    Image|endswith:
      - '\\powershell.exe'
      - '\\pwsh.exe'
    CommandLine|contains:
      - 'Invoke-WebRequest'
      - 'Invoke-RestMethod'
    CommandLine|contains|all:
      - 'http'
  filter_known_domains:
    CommandLine|contains:
      - 'microsoft.com'
      - 'windowsupdate.com'
  condition: selection and not filter_known_domains
falsepositives:
  - Legitimate DevOps/CI scripts fetching from newly onboarded but sanctioned third-party endpoints not yet in the allowlist
level: medium

Known false positives:

  • Legitimate automation/CI scripts fetching newly onboarded but sanctioned third-party endpoints not yet reflected in the domain allowlist

Tuning notes: Populate domain_first_seen.csv from registration event-stream / passive-DNS feeds and allowlisted_domains.csv from known brand/vendor domains; tighten first_seen_days_ago threshold once agent baselines are established.

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)
  • AI Agent/Coding-Assistant Runtime Downloading Archive from Unverified Lookalike Domain · unlock with TL Creds
  • Real-Time Storefront-Mirroring Scraper Process Against Brand Domain · unlock with TL Creds
  • Coordinated Dual-Domain Registration Beaconing to Shared Infrastructure Within Short Window · unlock with TL Creds
  • Brand-Clone Landing Page Triggers Out-of-Band APK Fetch · unlock with TL Creds
  • AI-Coding-Assistant-Generated Backend Artifact Deployed to Web Root Outside CI/CD · unlock with TL Creds
  • Financial Theft Indicator — High-Volume Payment/IBAN Field Exfiltration to External Endpoint · unlock with TL Creds
  • Web-Facing Backend Writing Combined Payment-Card + IBAN + National-ID Fields to a Single Log · unlock with TL Creds
  • Recurring Server-Side POST Beacon to Telegram Bot API (Non-Interactive C2 Pattern) · unlock with TL Creds

Indicators of compromise

As of 2026-06-30, Threadlinqs Intelligence tracks 21 indicators of compromise for TL-2026-1018.

Sample indicators (2 of 21)

  • File hash: eb07edaa2786cfddfa4c15526168f2200d85300aee0a8f253b32d2462a7b0bcd
  • File hash: 2202a30daad9928ef47cca5f4ab04ce083692a94428e386fa01c2dd44557e34b

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

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

Technical analysis

Unit 42 researchers (Keerthiraj Nagaraj, Diva-Oriane Marty, Beliz Kaleli, Oleksii Starov) systematically prompted two production LLMs — 'LLM1' (a production-optimized mini-class enterprise model, April 2025 build) and 'LLM2' (a low-latency lite-class frontier model, June 2025 build) — with 685,339 adversarial prompts referencing 913 global brands spanning Technology, Finance, Healthcare, E-commerce, Government, Gambling, and Logistics. The prompts, run across three temperature configurations (Precise T=0.1, Balanced T=0.7, Creative T=1.5), asked for customer support links, corporate portal logins, and software download pages. Because LLMs do not verify link existence before responding, they predicted plausible-looking URLs based on training-data patterns, generating 2.1 million unique URLs. Of these, 13,229 (0.61%) resolved to currently malicious infrastructure, 41,313 (1.90%) were high-risk parked/opportunistic domains, and 809,455 (37.28%) were non-existent domains (NXD) — roughly 250,000 of which remain unregistered and available for immediate adversarial preemption. LLM1 produced a 44.6% overall hallucination rate versus LLM2's 27.5%, though LLM2 skewed more toward higher-value subdomain- (45.1%) and domain-level (20.0%) hallucinations versus LLM1's path-level bias (56.6%), expanding the registerable attack surface. Confirmed malicious infrastructure broke down as 67.2% malware (drive-by downloads, exploit kits), 16.2% phishing (credential harvesting, brand impersonation), 13.7% grayware (adware/PUP installers), and 3.0% command-and-control. The technique creates a distinct supply-chain risk for autonomous AI agents, which can execute web requests to hallucinated domains without human verification; newly registered phantom domains carry zero reputation history, bypassing traditional URL/reputation filtering. Unit 42 documented multiple real-world exploitation windows: the 'Montana Empire' campaign (March 8-31, 2026, 23-day adversarial exploitation window/AEW) cloned a national postal service e-commerce marketplace storefront in real time, using an AI coding assistant to build a PHP-based phishing kit (7.96 MB ZIP archive, SHA256 eb07edaa2786cfddfa4c15526168f2200d85300aee0a8f253b32d2462a7b0bcd) that performed dual-channel payment interception (credit cards and IBAN transfers) plus national identity document harvesting, controlled through a Telegram-based C2 panel branded 'Kimseye Güvenme' ('Trust No One') supporting real-time OTP relay by a human operator. A second campaign (February 18-April 10, 2026, 51-day AEW) distributed a malicious Android APK (12.6 MB, SHA256 2202a30daad9928ef47cca5f4ab04ce083692a94428e386fa01c2dd44557e34b) impersonating a national postal delivery service app via a pixel-accurate brand-clone landing page with fabricated social-proof (4.8-star rating, '2M+ users'), delivered out-of-band outside official app marketplaces. Additional detections included a Bangladesh-targeted sports-betting credential harvester (45-day AEW), a second sports-betting lookalike registered 18 minutes after the first using identical infrastructure (40-day AEW) — indicating orchestrated regional targeting of Bengali-language markets with Bangladeshi Taka payment processing — a re-registered European retail-bank lookalike (35-day AEW), and an 11-month-old UAE commercial-bank lookalike historically validating corporate database-admin targeting. Unit 42 recommends proactive hallucination-surface mapping (simulating adversarial prompting against target LLMs to preemptively identify and monitor phantom domains), domain-registration event-stream monitoring, multi-signal verification (threat intel + content crawling + ownership analysis) before trusting AI-suggested URLs, and validating autonomous-agent web-request output before execution.

MITRE ATT&CK techniques by tactic:

Collection

T1005 Data from Local System

Credential Access

T1040 Network Sniffing; T1187 Forced Authentication

Command and Control

T1071 Application Layer Protocol; T1102 Web Service

Initial Access

T1199 Trusted Relationship; T1566 Phishing

Execution

T1204 User Execution

Resource Development

T1583 Acquire Infrastructure; T1587 Develop Capabilities; T1608 Stage Capabilities

Reconnaissance

T1589 Gather Victim Identity Information; T1590 Gather Victim Network Information; T1592 Gather Victim Host Information; T1598 Phishing for Information

Impact

T1657 Financial Theft

Remediation

Immediate actions

  • Enroll high-value brand domains and predictable AI-hallucinated variants in registration event-stream / newly-registered-domain monitoring so lookalikes are flagged at registration time rather than after weaponization
  • Deploy Advanced URL Filtering and Advanced DNS Security (or equivalent) to block access to newly registered, zero-reputation domains at the network egress layer
  • Block/alert on Telegram Bot API traffic (api.telegram.org) originating from unexpected endpoint or server-side processes, a common low-cost C2/exfil channel for AI-built phishing kits
  • Warn customers of high-value brands (banking, postal/logistics, sports betting) about out-of-band APK distribution and lookalike login pages during active AEW windows

Workarounds

  • Restrict autonomous AI agents from executing unauthenticated outbound web requests to domains outside an allowlist until multi-signal verification completes
  • Require human-in-the-loop confirmation before an AI agent acts on a URL it generated rather than retrieved from a verified source

Longer-term hardening

  • Run proactive hallucination-surface mapping: simulate adversarial prompting (varied temperature configurations) against the LLMs your own users/agents rely on to preemptively discover and pre-register or monitor brand-relevant phantom domains before attackers do
  • Deploy AI runtime security controls (e.g. Prisma AIRS or equivalent) and agentic endpoint security to validate any URL an autonomous AI agent is about to fetch before the request executes
  • Establish multi-signal verification (threat intelligence + content crawling + WHOIS/ownership analysis) as a gate before trusting any AI-suggested or AI-agent-generated URL in production workflows
  • Commission a formal AI security assessment of internally deployed LLMs/agents to quantify hallucination rate and structural hallucination pattern (path- vs subdomain- vs domain-level) as an attack-surface metric

Affected and fixed versions

  • Multiple (913 global brands) — Brand-facing customer portals, e-commerce marketplaces, postal/logistics services, banking portals, sports-betting platforms
    Vulnerable versions: Any brand referenced in LLM training data whose customer-facing URLs the model has not been grounded to verify
  • Unnamed LLM vendors — LLM1 (production-optimized mini-class enterprise model, April 2025 build); LLM2 (low-latency lite-class frontier model, June 2025 build)
    Vulnerable versions: LLM1 baseline (44.6% hallucination rate); LLM2 baseline (27.5% hallucination rate)

Timeline

  • LLM1 (production-optimized mini-class enterprise model) baseline used for Unit 42's adversarial hallucination study, exhibiting a 44.6% hallucination rate across the test corpus.
  • LLM2 (low-latency lite-class frontier model) baseline used for the study, exhibiting a 27.5% hallucination rate but a higher proportion of higher-value subdomain- and domain-level hallucinations.
  • Unit 42's discovery pipeline first identifies hallucinated URLs for the postal-delivery-app domain, marking the start of the 51-day adversarial exploitation window for the APK campaign.
  • Unit 42 detects the Montana Empire postal e-commerce hallucinated domain in its monitoring pipeline (13 distinct hallucinated URLs surfaced across both tested LLMs), starting the 23-day adversarial exploitation window.
  • Montana Empire campaign's documented adversarial exploitation window closes after 23 days of active credential and payment-data interception.
  • Adversary registers the Montana Empire lookalike domain and deploys the AI-assisted PHP phishing kit the same day, cloning a national postal service e-commerce marketplace storefront in real time.
  • Postal-service APK impersonation campaign's documented adversarial exploitation window closes after 51 days.
  • Unit 42 (Palo Alto Networks) publishes 'Phantom Squatting: AI-Hallucinated Domains as Supply Chain Threat,' documenting the 685,339-prompt, 913-brand study and associated real-world campaigns.

Sources cited

Previous and next in supply chain

Further reading

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