Large language models keep inventing web addresses that do not exist. Attackers have started buying those made-up domains before anyone else can, then hosting phishing pages on them to catch traffic that AI tools point their way.
Palo Alto Networks’ Unit 42 calls the trick phantom squatting, and its new research shows it is already happening in the wild.
The reason it matters is trust. Developers and AI assistants increasingly treat the links a model hands back as real. When a model invents a domain that does not exist yet, whoever registers it first inherits all of that misplaced trust, with no phishing email and no malicious ad required.
To measure the problem, Unit 42 asked two AI models 685,339 questions about 913 well-known brands across technology, finance, healthcare, government, gambling, and other sectors.
The models produced 2.1 million links. Threat intelligence already flagged 13,229 of them as outright malicious, meaning the AI was handing out known-bad addresses. Roughly 250,000 of the invented domains had no owner yet, each a ready target for whoever registers it first.
How phantom squatting works
The attack works because a brand-new domain has no reputation. Blocklists, threat feeds, and reputation scores all need a site to misbehave for a while before they flag it.
A freshly registered phantom domain has no such record, so those filters have nothing to flag. By the time they catch up, the victim has already been sent to the site by a tool they trust.
Two details make it worse. The fake domains were not sitting in the training data: both models shipped before the real malicious sites existed, so the addresses come from the models’ own language patterns, not memory. And those patterns are consistent.
Different models often invent the same fake domain for the same question, which makes an attacker’s next target easy to guess. Turning up a model’s “creativity” setting only produced more invented domains. As Unit 42’s researchers put it, the vector “exploits a structural property of LLM architectures that remains inherently unpatchable.”
Two observed cases
Two cases show the full loop. On March 8, 2026, Unit 42’s system predicted that AI models would invent a domain resembling a national postal service’s online marketplace. Both models generated it at every temperature setting, a strong sign that they treated the fake site as fact.
Twenty-three days later, on March 31, an attacker registered that exact domain and stood up a phishing kit named Montana Empire. The kit copied the real storefront in real time. It stole card numbers, bank-transfer details, and national ID data.
A Telegram bot lets the operator approve victims’ one-time passcodes by hand. The giveaway: leftover project files and session logs showed the criminal had built the kit with an AI coding assistant. Attacker and defender reached the same fake domain the same way, by asking an AI.
In the second case, Unit 42 flagged a hallucinated postal-service domain a full 51 days before an attacker registered it. The attacker then wrapped it in a pixel-perfect brand clone, added a fake 4.8-star rating and a claim of over two million users, and used it to push a malicious Android app.
Other detected domains impersonated a major UAE bank that an attacker had already been abusing for nearly a year, a European bank, and sports-betting sites aimed at users in Bangladesh.
An old trick with a new target
Phantom squatting is the domain version of slopsquatting, where attackers register the fake software package names that AI coding tools invent. That is not a hypothetical.
A large USENIX study found code-generating models routinely suggest package names that do not exist, and the PhantomRaven campaign turned exactly that behavior into malware hidden in 126 npm packages with more than 86,000 installs.
It points to a larger shift: model output is becoming input. Developers, agents, and security teams act on AI-generated links and names before anyone verifies them, and AI keeps shrinking the time defenders have to react.
It also lands in a world where brand-impersonation phishing is now a paid service, with kits like Lucid and Lighthouse standing up 17,500 fake domains against 316 brands in 74 countries.
What to do
Because models hallucinate consistently, security teams can map which fake domains a model is likely to produce and watch for anyone registering them, often with weeks of warning. For everyone else, the practical steps are simple:
- Do not trust a link just because an AI gave it. Confirm the domain is the real, official one before you type a password or paste it into code.
- Keep AI agents from automatically opening or downloading from model-generated links without a check. An agent has no instinct to hesitate the way a person might.
- Treat anything a model writes as an unverified draft, not an authority.
That window is open, and it rewards whoever moves first. The real question, as Unit 42 frames it, is simply whether defenders or attackers reach these domains sooner.
