A malicious custom GPT named “Plus 5.6” is using the legitimate chatgpt.com platform as an entry point for a ClickFix malware campaign, according to security firm Huntress. The GPT impersonates an official ChatGPT offering, claims capacity is limited, and directs users to a supposed alternate site to continue the conversation. The campaign does not rely on a vulnerability in ChatGPT itself; it exploits trust in user-created GPTs and search advertising.
The destination was a Google Sites page made to resemble a Cloudflare verification screen. Instead of performing a normal CAPTCHA check, it instructed visitors to copy and run a command in Windows. That is the core of ClickFix: convincing the target to execute the malware-delivery command themselves.
- The fake “Plus 5.6” GPT presents itself as an official-looking ChatGPT option and redirects users to an alternate address.
- The redirected page imitates a Cloudflare verification prompt and asks the visitor to execute clipboard content in Windows.
- An eight-stage infection chain then installs a remote-access trojan, using a legitimately signed Canon application and code concealed in an audio file.
Huntress examined at least 40 incidents connected to the malicious page, including two confirmed referrals from the fake GPT. The trojan described by the company can access a compromised system’s screen, camera, microphone and files, and can install additional malicious software.
A separate Island report documented a similar operation involving roughly 850 visits from paid ads, 26 ChatGPT-lookalike destinations and 71 Google Ads campaign identifiers. Those figures describe tracked infrastructure activity rather than the number of infected computers. That operation also used custom GPTs and shared chats to push an “overloaded service” narrative before directing users to a page that requested command execution; the resulting behavior was consistent with NetSupport RAT.
A CAPTCHA or other routine verification should never require opening PowerShell or the Windows Run dialog to paste a command. Anyone who has already run such a command should disconnect the device from the network, have it examined, and change important passwords from a known-safe device. The incident also shows why reaching a genuine platform domain is not, by itself, proof that instructions delivered by a user-created chatbot are safe.







