Phantom Squatting: The New Frontier of AI-Assisted Phishing

Has anyone dug into the Unit 42 report on "Phantom Squatting"? It’s a fascinating evolution in social engineering. Attackers are registering domains that LLMs "hallucinate" during conversations. When a user asks an AI for a link to a tool or documentation, the AI might invent a URL, and if an actor has already squatting on it, the user gets served malware or a phishing kit.

Since LLMs often generate plausible-looking but non-existent domains, these domains are fresh and usually lack reputational flags. This bypasses traditional security layers that rely on blocklists or long-standing reputation scores. It turns the AI's tendency to confabulate into a delivery mechanism for payloads.

We should probably start monitoring for traffic originating from known AI service user-agents or referrers heading to low-reputation domains. Here’s a quick Python snippet to help identify potential risk based on domain registration age, a key indicator of these squatting domains:

import whois
import datetime

def check_domain_risk(domain):
    try:
        w = whois.whois(domain)
        creation_date = w.creation_date
        if isinstance(creation_date, list):
            creation_date = creation_date[0]
        
        if creation_date:
            age_days = (datetime.datetime.now() - creation_date).days
            if age_days < 30:
                return f"[!] High Risk: Domain is only {age_days} days old."
            return f"[+] Safe: Domain is {age_days} days old."
    except Exception as e:
        return f"[?] Unknown/Unregistered: {str(e)}"

# Example usage
print(check_domain_risk("example-hallucinated-domain.com"))


I'm curious: Are any of you seeing policies emerging around restricting internal tool downloads to verified repositories to mitigate this specific vector?