Attackers exploit LLMs that hallucinate brand domains to create plausible yet unregistered portals, API endpoints, or services that misdirect users or automated agents.

Defensive guidance points to Palo Alto Networks’ suite and Unit 42 AI Security Assessment for safe AI use and development.

Unit 42 analyzed 913 brands and found about 250,000 hallucinated domains and more than 13,000 confirmed malicious URLs linked to brands.

AI prompts can reveal official website domains or downloads, enabling the generation of fake domains through AI hallucinations.

Mitigations include verifying URLs against authoritative docs or allowlists, restricting AI agents from freely connecting to new domains, and tightly controlling credentials and data access.

A proactive discovery framework was built with a query agent and a URL creator agent across two LLM families and multiple temperatures to map hallucination behavior and risk.

The broader implication is that LLMs become trusted supply-chain dependencies whose outputs can be weaponized if not independently verified, necessitating proactive monitoring and rapid registration alerts.

Defensive advice emphasizes identifying likely hallucinated domains and registering them preemptively, while monitoring domain-registration streams to respond quickly.

Phantom squatting is an emerging AI-driven threat where attackers register fictitious, brand-aligned domains that can intercept traffic steered by AI systems and phishing tools.

In large-scale probing, researchers generated 2.1 million URLs from 685,339 prompts across two LLMs and temperatures, identifying 13,229 malicious URLs and roughly 250,000 unregistered domains.

The findings underscore opportunities for attackers to preemptively register phantom domains and exploit the software supply chain.

This phenomenon expands risks like slopsquatting and typosquatting, driven by the probabilistic nature of large language models.