A fake ChatGPT on the real chatgpt.com told people to paste a command
Someone searched Google for “chatgpt.” They clicked the top result, which was a paid ad. It took them to chatgpt.com — OpenAI’s real website, with the real address in the browser bar. The chatbot that opened told them the service was having trouble and offered a backup page. On that page, a “prove you’re human” box asked them to copy a command and paste it into Windows. They did. It installed a remote access trojan: software that hands control of the computer to a stranger.
The security firm Huntress published the full chain on September 28, 2026, in a report titled “Attackers Abuse ChatGPT Custom GPTs to Deliver RAT via ClickFix.” Every number and detail below comes from that report.
This one deserves a few minutes of your attention even if nobody at your company thinks about security, because of how ordinary the victim’s behavior was. They did not open a strange attachment. They did not click a link in a suspicious email. They searched for a tool their company already uses, took the first result, and ended up on the correct website. Everything they were ever taught to check came back clean.
First, the thing that did not happen: OpenAI was not hacked
Nobody broke into OpenAI. No system was breached, and no account data was stolen. What the attacker used was a normal, advertised feature.
ChatGPT lets anyone publish a Custom GPT — a chatbot you set up yourself by writing instructions for it, which anyone can then open. Think of it as a page you publish on someone else’s website. And because ChatGPT hosts those chatbots, a published Custom GPT lives at an address that starts with chatgpt.com. That is not a flaw. It is how the feature works, the same way a video you upload lives on a youtube.com address.
The attacker named their chatbot “Plus 5.6” so it would read like a real ChatGPT model, published it, and bought Google ads pointing at it. Huntress found two instances of it. OpenAI removed the first one by September 25, 2026. A replacement tied to the same campaign was up two days later, on September 27.
How it actually worked, in order
- The search. A person searches for “chatgpt.” The top of the page is a sponsored result — an ad, sitting above the real listings. Huntress recovered the ad-tracking parameters still attached to the link, which is how they know the traffic was paid.
- The landing. The ad sends them to chatgpt.com. The domain is genuine, because the chatbot really is published there.
- The excuse. “Plus 5.6” answers with a “Service Availability Notice,” telling the visitor the service is “experiencing limited availability on the primary domain.” It offers two ways out: upgrade to Plus, or keep working through a “backup domain.”
- The handoff. The backup domain is a page on Google Sites — Google’s free website builder — dressed up with both ChatGPT’s branding and Cloudflare’s. Cloudflare is the company behind many of the “checking your browser” screens you see online, so the page reads as a routine security check. (The address was
sites[.]google[.]com/view/antibot172881, written here with brackets so it is not clickable. Do not visit it.) - The ask. The fake check tells the visitor to copy a command and paste it into Windows. That is the entire trick. Security researchers call it ClickFix: instead of getting you to download a file your antivirus might stop, it gets you to type the attack in yourself.
- The install. The pasted command quietly downloads an installer that presents itself as “Advanced Printer Configuration Reader.” It drops a real, legitimately signed Canon program on the machine and tucks a tampered file next to it, so the trusted program loads the attacker’s code instead of its own. A later stage hides its payload inside what looks like a WAV audio file. The end result is a remote access trojan — someone else operating the computer, with the signed-in user’s access to email, files, and saved passwords.
How many businesses got hit
Huntress says its security operations center responded to at least 40 incidents stemming from that one Google Sites page. Of those, two were confirmed to have arrived by way of the fake Custom GPT.
That distinction is worth keeping straight, because it will get blurred as this story travels. Forty-plus is the count of incidents traced to the fake “prove you’re human” page, which attackers were feeding from more than one direction. Two is the count confirmed to have come through the chatbot on chatgpt.com specifically. Both numbers are small. Neither is the point.
The point is the delivery route, and the route is cheap to rebuild. Publishing a Custom GPT is free. A Google Sites page is free. The only real cost was the ads, and anyone can buy those. That is why the first takedown bought two days of quiet.
Every brand the victim checked was real
Go back through the chain and look at what an alert employee would have verified.
- Check the address bar? It said chatgpt.com. Correct.
- Recognize the security check? It looked like Cloudflare, which is a real and extremely common security service.
- Make sure the software is signed by a real company? It was — Canon.
- Arrive through a trustworthy path? Google search, top result.
Four checks, four passes. This is the uncomfortable part of the story for anyone who has run a security awareness session: the advice we give people — look at the URL, look for the padlock, look for the publisher — did nothing here. Not because the advice is wrong, but because the attacker simply stopped tripping those particular wires.
Exactly one thing in the whole sequence was abnormal: being asked to copy a command and paste it into the computer.
What to do about it
You do not need a security program to shut this down. You need one sentence that everyone on your payroll knows.
No legitimate website, security check, or software update ever asks you to copy a command and paste it into Windows PowerShell, the Mac Terminal, or the Run box. That request, by itself, is the attack. Nothing else about the page has to look wrong, and in this campaign nothing else did.
Beyond that sentence, a few things make the next version of this easier to survive:
- Tell your team the rule out loud, by name. Say “ClickFix” and describe the paste box. People remember a shape better than a policy.
- Reach your tools by bookmark, not by search. Bookmark chatgpt.com and the other services your staff use daily. Searching a brand name and taking the top result means trusting whoever paid the most for that slot today.
- Make “I pasted something weird” a safe thing to report. The gap between the paste and the damage is minutes. An employee who is afraid of looking foolish will wait a day, and a day is plenty.
- Know who to call before you need them. If a machine ran one of these commands, it should be disconnected and examined by someone who does this for a living, and the passwords saved on it should be treated as known to strangers.
- Remember that “it was on the official site” is no longer proof of anything when the official site lets the public publish. That applies well beyond ChatGPT.
The broader shift
AI tools arrived at most companies from the bottom up. Nobody rolled out ChatGPT; employees started using it, and they got to it the way they get to everything — by searching for it. That habit is now a documented attack surface, and the content they land on can be written by anyone, hosted by the vendor, and sold to them through an ad.
None of this requires your business to be interesting to criminals. It requires one person, one search, and one paste.
We break down stories like this one for owners and managers who have to make decisions about AI without a security team behind them — more of it is in the AIWAS blog, and short versions land on @AIWASai.
Stop leaving customers on read.
We answer your reviews, keep your social pages posting, and make your ads convert, so you get more customers while you run the business.
Get Started