Executive Summary
In late September 2026, researchers identified a campaign where attackers created Custom GPTs on the legitimate ChatGPT.com domain to impersonate new AI models (e.g., 'Plus 5.6'). These GPTs directed users to a malicious 'backup domain' hosted on Google Sites. This site functioned as a ClickFix lure, tricking users into pasting PowerShell commands that initiated an eight-stage infection chain.
The technical complexity of this campaign is significant, utilizing multi-stage obfuscation, steganography (hiding payloads in .wav and NuGet files), and DLL sideloading via legitimate Canon and Stardock binaries. The final payload is a highly functional RAT capable of screen broadcasting, audio capture, and further malware deployment. Over 40 incidents have been confirmed, demonstrating the success of abusing trusted AI platforms to bypass traditional security filters.
This campaign highlights an evolving trend where attackers leverage the perceived trust of AI platforms like OpenAI and Anthropic. The rapid rotation of lures—switching from Canon-signed binaries to Stardock-signed ones—suggests a maintained, professional malware framework designed for longevity and evasion.
