Skip to main content
All incident reports
Incident report

Fraudulent ChatGPT Custom GPTs used to push ClickFix malware, Huntress finds

Attackers built custom versions of ChatGPT that posed as a product called Plus 5.6 and sent users to a fake verification page that tricks them into running PowerShell. Huntress says the chain ends in a remote access trojan; OpenAI removed the first custom GPT on 25 September.

· First published

The short version

  • The lure. A threat actor published Custom GPTs — user-built variants of ChatGPT hosted on chatgpt.com — named Plus 5.6 and promoted them through sponsored Google results for searches such as "chatgpt". The GPT told users the main service was short of capacity and pointed them to a backup domain.
  • The payload. The backup page, hosted on Google Sites, shows a fake Cloudflare check and instructs visitors to run a PowerShell command. That command installs an MSI package that sideloads a malicious DLL through a legitimate Canon-signed application and ends in a remote access trojan (RAT) with remote desktop, camera, microphone and file-search capability.
  • The scale. Huntress says it investigated at least 40 incidents tied to the Google Sites page but confirmed only two that involved a Custom GPT. SecurityWeek and The Hacker News report at least 40 users infected in the campaign as a whole.
  • The takedowns. OpenAI removed the first Custom GPT by 25 September. Researchers found a second on 27 September, still active when they published. BleepingComputer reports OpenAI plans to retire custom GPTs on 11 December.

Timeline

DateEvent
25 SeptOpenAI takes down the first malicious Custom GPT.
27 SeptHuntress finds a second Custom GPT running the same campaign; it is still active at publication.
29 SeptHuntress's findings are reported publicly.
11 DecOpenAI is due to retire custom GPTs, according to BleepingComputer.

How the attack works

Victims arrived through a sponsored Google result and landed on a Custom GPT page on the genuine chatgpt.com domain, which lent the operation credibility. The GPT, titled Plus 5.6 and presented as the work of a community builder, replied to prompts with a link to what it called a backup domain. A notice told users to upgrade or switch to that domain for immediate access.

The backup domain was a Google Sites page carrying a fake Cloudflare CAPTCHA. This is a ClickFix lure: the page tells the visitor to paste a command that will supposedly prove they are human. The command runs PowerShell, which fetches an MSI installer. The installer abuses a legitimate Canon-signed executable to sideload an altered Canon DLL, which in turn loads an unsigned DLL, and the loader is hidden inside a .WAV audio file.

The final stages unpack a persistence script and the RAT from a custom encrypted archive; Huntress researchers described it as "basically a homemade, encrypted zip file". Persistence comes from a registry Run key and a scheduled task, both named Canon Configuration Reader. Later versions of the kit swapped the Canon binary for a Stardock-signed one and hid the loader in a NuGet package, but the payload was unchanged.

What the malware does

According to Huntress, the RAT can open remote desktop sessions, capture the screen, camera, microphone and system audio, search file contents, list installed antivirus and Defender status, recognise 17 web browsers and launch the default one, and drop further executables, DLLs and scripts. Before running it checks for virtual machine artefacts and bypasses user-mode monitoring tools.

The implant locates its command-and-control server using DNS-over-HTTPS — domain lookups wrapped in ordinary encrypted web traffic — through Cloudflare, Google and Quad9 resolvers. Huntress notes that these lookups therefore do not appear in local DNS logs. Separately, the researchers point to process-level signs, such as PowerShell silently launching an MSI installer from the temporary folder.

Who was affected

The sources describe the numbers differently. SecurityWeek says at least 40 users were infected in the campaign, with at least two incidents tied to a Custom GPT. The Hacker News says no less than 40 users were infected. BleepingComputer reports that Huntress investigated at least 40 incidents connected to the Google Sites page but confirmed only two involving a custom GPT variant. The sources do not identify any victim organisation.

Huntress says abusing trusted platforms this way is a pattern rather than a one-off, and The Hacker News places the activity alongside other ClickFix campaigns that use Google Sites pages mimicking OpenAI Codex and Anthropic Claude. No source states whether those other campaigns share an operator with this one.

What is not known yet

  • Who is behind the campaign, and whether they also ran the earlier ClickFix lures that used shared ChatGPT conversations.
  • Whether the second Custom GPT was removed, and when.
  • The true number of infected victims, and whether the figure of at least 40 covers only the incidents tied to the Google Sites page.
  • What data, if any, the RAT took from infected machines.
  • Whether the sponsored Google adverts that led users to the GPTs have been taken down.

What to do

If you use ChatGPT

  1. Reach ChatGPT by typing chatgpt.com yourself rather than clicking sponsored search results.
  2. Close any GPT that claims limited capacity on the main service and pushes you to a backup domain.
  3. Never paste a command from a CAPTCHA or verification page into PowerShell or the Windows Run box.

If you run Windows endpoints

  1. Alert on PowerShell spawning msiexec.exe to install an MSI from the temporary folder.
  2. Hunt for signed applications launching from unexpected subfolders under %LOCALAPPDATA%\Programs\.
  3. Look for a Run registry value and a scheduled task named Canon Configuration Reader that reappear after deletion.
  4. Check for non-browser processes resolving names over DNS-over-HTTPS to public resolvers.

For everyone else

  1. Treat any web page that asks you to run a command as hostile, however convincing the verification step in front of it looks.

Update log

  • First published.

Sources

New incident reports by email

One email a week with the reports published since the last one.

Want us to look at your domain?

Give us one apex domain and we'll send back a written snapshot of what is already visible from outside. No call required.