Custom GPTs can be useful. You can find GPTs designed for coding, writing, research, education, productivity, data analysis, and many other tasks.
But there is an important security question that more users should be asking:
Can a Custom GPT be malicious even when it appears on the real ChatGPT website?
Yes.
A Custom GPT is not automatically trustworthy simply because its page is hosted on chatgpt.com.
This became a very practical concern in September 2026, when security researchers documented a campaign in which attackers created malicious Custom GPTs and used them as the first stage of a malware-delivery chain. The malicious GPTs were hosted on the legitimate ChatGPT domain and were designed to send users toward a fake verification page that eventually attempted to install remote-access malware.
The incident highlights a broader security lesson:
A trusted platform can host user-created content that is not trustworthy.
That does not mean Custom GPTs are inherently dangerous.
It means users need to understand the difference between the platform and the content created by users on that platform.
What Is a Custom GPT?
A Custom GPT is a customized version of ChatGPT created for a particular purpose.
Depending on the platform and configuration, a creator can provide a Custom GPT with things such as:
- custom instructions
- specialized behavior
- knowledge files
- specific workflows
- available tools or capabilities
- connections to external services in supported configurations
That makes Custom GPTs useful for specialized tasks.
For example, someone might create a GPT that helps explain programming concepts or summarize documents.
Another creator might build one for marketing, mathematics, cybersecurity, language learning, or business analysis.
The important point is that the creator controls the GPT's configuration.
When you use a GPT created by someone else, you are interacting with something that another person configured.
Can a Custom GPT Be Malicious?
A Custom GPT can be abused as part of a malicious workflow.
That does not necessarily mean the GPT itself is exploiting a vulnerability in ChatGPT.
In the September 2026 campaign investigated by Huntress, attackers created Custom GPTs with names designed to look like legitimate ChatGPT offerings.
One malicious GPT was called “Plus 5.6.”
Victims could encounter it through sponsored Google search results while looking for ChatGPT.
The GPT was hosted on the legitimate ChatGPT website.
When users interacted with it, the GPT displayed a fake service-availability message and directed them toward a supposed backup website.
The external site then presented a fake Cloudflare-style verification page.
Instead of performing a normal CAPTCHA, the page instructed the victim to copy and execute a command.
That command started a malware-delivery chain that eventually deployed a remote-access trojan.
Huntress investigated at least 40 related incidents and confirmed two incidents that originated from the malicious Custom GPT lure.
How Can a Malicious GPT Be on the Real ChatGPT Website?
This is the part that confuses many users.
You may have learned:
“Check the domain. If it says chatgpt.com, it is safe.”
That rule is too simple.
The correct mental model is:
Trusted platform does not automatically mean trusted user-created content.
Consider a hypothetical example.
You visit:
https://chatgpt.com/
The domain is legitimate.
You open a Custom GPT created by another user.
The page is still legitimately hosted by ChatGPT.
But the content and instructions of that Custom GPT may have been created by an unrelated person.
The domain is real.
The creator may still be untrusted.
This is similar to other platforms where users can publish content under a trusted domain.
A legitimate marketplace can contain a malicious application.
A legitimate social network can contain a malicious post.
A legitimate file-sharing service can contain a malicious file.
And a legitimate AI platform can contain user-created content that is designed to manipulate users.
The Most Important Red Flag
If a Custom GPT tells you to leave ChatGPT and perform a dangerous action on your computer, stop.
Especially if it asks you to:
- open Windows Run
- open PowerShell
- open Command Prompt
- open Terminal
- paste a command
- download an unknown installer
- disable antivirus protection
- disable Windows Defender
- install a browser extension from an unknown source
- visit an unfamiliar “backup” website
A chatbot should not need you to execute an unknown operating-system command just to continue a normal conversation.
This is one of the strongest practical warning signs.
The “Backup Website” Trick
The recent malicious Custom GPT campaign used an especially effective social-engineering pattern.
The GPT claimed that the normal service was experiencing limited availability.
It then offered the user a supposed alternative or backup domain.
This creates a believable explanation for why the user is leaving the trusted platform.
The sequence looks something like this:
Google Search
|
v
Sponsored result
|
v
Real chatgpt.com
|
v
Malicious Custom GPT
|
| "Service unavailable"
v
Fake backup website
|
v
Fake verification page
|
v
ClickFix command
|
v
Malware
Notice the psychological trick.
The attacker does not need to make the first website look fake.
The first step can be the real ChatGPT website.
The malicious behavior begins with the user-created content and the external destination.
Why “It Was on ChatGPT.com” Is Not Enough
Domain reputation is useful, but it is only one security signal.
Imagine three different situations.
| Situation | What you know | Risk question |
|---|---|---|
| Official ChatGPT conversation | You are interacting directly with the core service | What information are you sharing? |
| Third-party Custom GPT | A user created the configuration | Who created it and what does it ask you to do? |
| External website linked by a GPT | You have left the original platform | Is the destination legitimate and necessary? |
Each step introduces a different trust boundary.
What Information Should You Avoid Giving to an Unknown Custom GPT?
Even when there is no malware involved, you should be careful about the information you provide to third-party AI tools.
Avoid sharing highly sensitive information with an unfamiliar Custom GPT unless you have a clear reason to trust its creator and understand the applicable data handling.
Examples include:
- passwords
- one-time authentication codes
- private API keys
- recovery codes
- cryptocurrency seed phrases
- banking credentials
- government identity documents
- private company secrets
- confidential customer information
- unpublished source code containing secrets
A good security rule is:
If you would not paste the information into an unknown third-party website, do not casually give it to an unknown third-party GPT.
Can a Custom GPT Steal Your Password?
A Custom GPT should not be treated as a trusted password manager or authentication system.
Never give a GPT your password simply because it asks for one.
The same applies to:
- MFA codes
- recovery codes
- private keys
- session tokens
- API secrets
There is almost never a legitimate reason for a chatbot to need your actual authentication secret.
If a GPT says:
“Send me your password so I can verify your account.”
stop immediately.
That is not normal account verification.
What About Uploaded Files?
Files deserve the same caution as passwords.
If you upload a document to an unfamiliar GPT, consider what information is contained in it.
A document might include:
- personal information
- customer records
- financial information
- internal company information
- source code
- API credentials
- internal URLs
- security configuration
Do not assume that a GPT created by an unknown person should be trusted with confidential files.
How to Check Who Created a Custom GPT
Before trusting a Custom GPT, inspect its creator information where available.
Ask:
- Who created it?
- Is the creator identifiable?
- Is the creator associated with a real organization?
- Does the organization have an official website?
- Does the GPT have a clear purpose?
- Does its behavior match its stated purpose?
Be particularly careful with names that appear to imitate official products.
A name such as:
ChatGPT Plus 5.6 Official
does not automatically mean that it was created by OpenAI.
Attackers deliberately choose names that exploit current product names, model releases, subscriptions, and popular searches.
Do Reviews Make a Custom GPT Safe?
No.
Reviews can be useful, but they should not be treated as a security guarantee.
Reviews can be fake.
They can also be outdated.
A previously harmless GPT can change behavior if its creator changes its configuration.
Therefore, evaluate the actual behavior of the GPT rather than relying entirely on its popularity.
Why Attackers Are Interested in AI Platforms
AI platforms have become extremely familiar to internet users.
People increasingly use them for:
- research
- coding
- writing
- education
- productivity
- data analysis
- business workflows
That creates a powerful social-engineering opportunity.
If an attacker can make malicious instructions appear inside a platform the victim already trusts, the victim may be less suspicious.
This is the same basic psychological principle behind many successful phishing attacks:
Borrow trust from a familiar brand.
The difference is that AI platforms can provide an interactive interface rather than a simple static phishing page.
The 2026 Custom GPT Malware Campaign
The September 2026 campaign is a useful example of how this works in practice.
Attackers created a Custom GPT called “Plus 5.6” and promoted it through sponsored search results for ChatGPT-related searches.
The malicious GPT was hosted on the legitimate ChatGPT domain.
When users interacted with it, the GPT returned a scripted service-availability message.
The message directed victims to a supposed backup site.
The external page displayed a fake Cloudflare verification experience.
The victim was then instructed to perform a ClickFix-style action involving a command.
Running the command began an infection chain that eventually deployed a remote-access trojan.
Huntress reported that the malware could collect extensive information from an infected Windows system and provide attackers with capabilities such as remote access and additional payload execution.
The important lesson is not the specific malware family.
The important lesson is the trust chain:
Google → ChatGPT → Custom GPT → Google-hosted page → fake Cloudflare → malware.
Several trusted brands were used in sequence.
Does This Mean You Should Stop Using Custom GPTs?
No.
That would be an unnecessary conclusion.
Custom GPTs can be useful tools.
The correct approach is to use them with appropriate trust boundaries.
Think of an unfamiliar Custom GPT like an unfamiliar browser extension or third-party application.
You do not automatically assume it is malicious.
But you also do not automatically give it sensitive information or follow instructions that affect your operating system.
Five Rules for Safer Custom GPT Use
Rule 1: Never share authentication secrets
Do not give a GPT passwords, MFA codes, recovery codes, API keys, or private keys.
Rule 2: Treat external links cautiously
If a GPT suddenly tells you to leave ChatGPT and visit another website, stop and evaluate why.
Rule 3: Never run commands from a GPT
If an AI tool tells you to paste a command into PowerShell, Command Prompt, Terminal, or Windows Run, do not execute it blindly.
Rule 4: Verify the creator
Prefer GPTs from creators or organizations you can independently identify and trust.
Rule 5: Use a separate security mindset for files
Do not upload confidential documents to unfamiliar GPTs simply because the GPT claims to perform a useful analysis.
What If a Custom GPT Tells You to “Verify You Are Human”?
Be extremely cautious.
A normal CAPTCHA may ask you to:
- click a checkbox
- select images
- solve a visual challenge
A CAPTCHA should not require you to:
- open PowerShell
- open Command Prompt
- press Windows + R and paste a command
- disable security software
- download an unknown executable
- execute code to “prove you are human”
If you see such a request, close the page.
This is one of the clearest signs of a ClickFix-style attack.
What If You Already Followed a Suspicious GPT's Instructions?
The response depends on what you did.
If you only chatted with the GPT
Review what information you shared.
If you provided sensitive information, take appropriate steps to protect it.
If you clicked an external link
Close the page and do not download or execute anything it requests.
If you downloaded a file
Do not open it.
Delete or quarantine it and scan the system using reputable security software.
If you executed a command
Treat the device as potentially compromised.
Disconnect it from the network if appropriate and use a separate trusted device to secure important accounts.
Review active sessions and change exposed credentials.
If it is a work computer, contact your organization's security or IT team immediately.
Why a Password Reset May Not Be Enough After Malware
If malicious software executed on your computer, the attacker may have attempted to steal more than your password.
Depending on the malware, attackers may target:
- browser passwords
- authentication cookies
- session tokens
- saved payment data
- browser autofill information
- cryptocurrency wallets
- local documents
That is why a suspected malware infection should not be treated as a simple password-reset event.
Secure the affected accounts, revoke sessions where possible, investigate the computer, and determine whether additional credentials were exposed.
The Bigger Lesson: Trust the Behavior, Not Just the Domain
Security advice often tells people to look at the URL.
That remains useful.
But modern attacks increasingly exploit legitimate services.
An attacker may use:
- a legitimate social network
- a legitimate cloud-storage service
- a legitimate advertising platform
- a legitimate AI platform
- a legitimate code-hosting service
The infrastructure can be real while the content is malicious.
Therefore, security decisions should also consider what the content is asking you to do.
If a Custom GPT suddenly asks you to install software, execute a command, disable security controls, or enter an authentication secret, the behavior itself is a warning sign.
Quick Custom GPT Safety Checklist
- Check who created the GPT.
- Do not assume
chatgpt.commeans every user-created GPT is official. - Do not share passwords or MFA codes.
- Do not upload confidential documents to unknown GPTs.
- Be cautious about external links.
- Never run commands provided by a GPT without independently verifying them.
- Never disable antivirus or security controls because a GPT asks you to.
- Be suspicious of fake service-outage or “backup domain” explanations.
- Treat unexpected CAPTCHA instructions involving Windows commands as a major red flag.
- Prefer verified creators and organizations for sensitive tasks.
Bottom Line
Custom GPTs are not inherently malicious.
But they should not automatically be treated as trusted simply because they are hosted on the real ChatGPT website.
The September 2026 “Plus 5.6” campaign demonstrated why. Attackers were able to use a legitimate ChatGPT-hosted Custom GPT as the first stage of a malware campaign without needing to compromise the ChatGPT domain itself.
The most important distinction is:
Trust the platform separately from the content created on the platform.
If an unfamiliar Custom GPT asks for your password, MFA code, private key, confidential document, software installation, or operating-system command, stop.
Do not follow the instruction simply because the GPT appears inside a legitimate AI service.
And if a supposed AI assistant tells you that you must open PowerShell, Windows Run, Terminal, or Command Prompt and paste something to “verify,” “fix,” or “activate” the service, treat that as a serious malware warning.
The safest approach is simple:
Use AI tools for information and assistance. Do not let an unfamiliar AI tool become the authority that tells you what executable code to run on your computer.