Attackers used fake custom ChatGPTs to trick users into installing remote-control malware


Security firm Huntress says a fake custom ChatGPT called 'Plus 5.6', promoted through sponsored Google results, sent people to a fake 'security check' page. That page told them to run a command that installs malware able to take control of a Windows computer.
Attackers have been using OpenAI's own custom-GPT feature to lead people toward malware. That is the finding of Huntress, a managed detection and response (MDR) company, which means it monitors customers' systems for attacks and helps respond to them. BleepingComputer reported the research on September 29, 2026. Huntress researchers say the campaign affected dozens of users.
A custom GPT is a version of ChatGPT that someone builds for a specific task. It combines its own instructions, extra knowledge and skills. OpenAI hosts these custom GPTs, and their creators can publish them for other people to install and use. In this case, the attacker named the malicious GPT 'Plus 5.6'. The GPT was promoted in sponsored Google search results.
The fake GPT was set up to send users to a supposed backup site hosted on Google Sites. That page shows a fake Cloudflare check, imitating the 'verify you are human' screens many websites use. It then tells visitors to run a PowerShell command. PowerShell is a built-in Windows tool for running commands. Tricking people into pasting and running a command like this is known as a ClickFix attack.
If someone runs the command on their computer, it installs a malicious MSI, a Windows installer package. The installer launches a legitimate, digitally signed application alongside a modified DLL, a code library that the application loads. That modified library then loads the malware. The final malware is a remote access trojan (RAT), which is hidden software that gives an attacker remote control of a computer. According to Huntress, this one can take over the remote desktop, record audio and camera, search files, gather information about the computer and run additional malicious programs.

KrebsOnSecurity identifies the arrested Canadian as Edward Dubrovsky and reports extortion-related charges, a sealed complaint and a case transfer to Texas.
To keep running after a restart, the malware adds an entry to the Windows Registry's Run key and creates a scheduled task. Both are named 'Canon Configuration Reader'.
Analysis
Huntress had already seen similar attacks that used deceptive ChatGPT conversations to launch ClickFix tricks. It says using custom GPTs is a new approach. In both kinds of attack, the malicious instructions are hosted on the real ChatGPT.com domain. According to the report, this makes the operation look legitimate and raises the chance that victims will follow the instructions. A user who checks the web address would see a genuine OpenAI site, not an obvious fake.
The size of the custom-GPT part of the campaign is still unclear. Huntress says it investigated at least 40 incidents connected to the Google Sites page. It confirmed that only two of them involved a custom GPT, so the 40 incidents should not be read as 40 confirmed victims of fake GPTs.
OpenAI took down the first malicious GPT by September 25. Two days later, on September 27, the researchers found a second GPT linked to the same campaign. That GPT was still active when Huntress published its report. Separately, BleepingComputer reports that OpenAI plans to retire custom GPTs altogether on December 11.
Newer versions of the attack switched from an application signed by Canon to one signed by Stardock. They also changed how the loader, the code that starts the malware, is hidden and delivered. The malware itself stayed the same.
Huntress singled out one stage of the attack. The attackers built their own encrypted file system to hide the RAT and the script that keeps it running. "Instead of one encrypted blob, it's a custom archive with its own folder tree, basically a homemade, encrypted zip file," the researchers say. They describe a small header, followed by an index of 1,128 entries and then the file contents.
Huntress says most of the infection runs in memory or relies on files that look harmless. For that reason, it recommends that defenders watch what programs are doing, rather than scan files alone. The warning signs it lists include:
9 min

Browser extensions and tools that add code to Microsoft's sign-in page will stop working there. Users will still be able to sign in, and apps that authenticate through Microsoft's library or APIs are not affected.
3 min

Parents can choose who can add their teen to groups or see their profile photo, and see when the teen joins or leaves a group. They cannot read messages. One analyst says making the controls optional is an obvious weakness.
3 min
Correction ·
Corrected spelling and punctuation in the Armenian text.