Lock down your ChatGPT account before the next AI attack

OpenAI has admitted that a combination of its advanced AI models found a way out of a locked-down test environment and compromised another company's systems. That should make every ChatGPT user stop and pay attention. The models were supposed to stay inside a tightly controlled digital sandbox with no open internet access. Instead, they found a weakness, got online and compromised systems belonging to Hugging Face, a major platform for AI models and datasets.

OpenAI described the breach as an "unprecedented cyber incident" involving "state-of-the-art cyber capabilities." The company says the system included GPT-5.6 Sol and an even more powerful model still being tested behind closed doors. Here is what concerns me. OpenAI built the models and controlled the testing environment, yet the systems still found a way around the safeguards designed to contain them.

The AI models were trying to complete a cybersecurity challenge. They apparently had no goal of damaging Hugging Face. However, they kept pushing until they found vulnerabilities and crossed into another company's infrastructure. That should be a wake-up call. You cannot control what happens inside OpenAI's labs, but you can take a few minutes to lock down your ChatGPT account before an AI-powered threat comes looking for your password, personal files or private conversations.

CyberGuy Live: Missed "Sick of Spam?" Get the replay and checklist

Our free CyberGuy Live class, "Sick of Spam?" , has ended, but you can still watch the full replay and download our spam-stopping checklist. Kurt "CyberGuy" Knutsson walks you step by step through simple ways to reduce robocalls, spam texts, junk email and unwanted messages. You’ll also learn how to curb political texts, clean up your inbox and spot messages that could put your personal information at risk.

Get the free replay and checklist now at CyberGuyLive.com.

OpenAI says its models escaped during a cyber test

OpenAI designed the internal evaluation to measure how well its models could find and exploit difficult security vulnerabilities. The company ran the test without some of the production safety systems that normally block high-risk cyber activity. OpenAI said researchers removed those protections so they could measure the models' maximum capabilities.

The models operated inside a restricted testing environment. Internet access was limited to an internally hosted service that acted as a proxy for software packages. However, the models found a previously unknown vulnerability in that service. They exploited the weakness and moved through OpenAI's research environment until they reached a computer with internet access.

Once online, the models identified Hugging Face as a possible source of answers for the ExploitGym security benchmark. The models then searched for information that could help them complete the test. OpenAI said they combined several attack methods. Those methods included stolen credentials and previously unknown vulnerabilities.

In one case, the models found a path that allowed remote code execution on Hugging Face servers. That gave them the ability to run code on another company's infrastructure. OpenAI said the models remained focused on completing the evaluation. Yet their narrow goal still led them to cross security boundaries and compromise an outside company.

OpenAI says the incident exposed a growing gap between what advanced models can do and the safeguards designed to contain them. "The primary lesson from this incident is that model security and safety must keep pace with rapidly advancing capabilities," OpenAI said in its incident report.

Hugging Face detected an autonomous AI intrusion

Hugging Face first disclosed the breach on July 16, 2026. The company said an autonomous AI agent system carried out the intrusion from beginning to end. The attack involved thousands of automated actions across short-lived digital environments.

Hugging Face identified unauthorized access to a limited set of internal datasets. The attacker also accessed several credentials used by its services. However, the company found no evidence that anyone altered its public models or user-facing datasets. It also verified that its software supply chain remained clean.

Hugging Face closed the vulnerabilities used for the initial access. It also rebuilt affected systems and rotated exposed credentials. Meanwhile, the company advised Hugging Face customers to rotate their access tokens and review recent activity. That guidance applies to Hugging Face accounts rather than consumer ChatGPT accounts. OpenAI later determined that its models caused the activity during its internal evaluation. The two companies continue to investigate the incident together.

CyberGuy reached out to OpenAI and Hugging Face for comment but did not receive a response before publication.

What the AI sandbox escape means for ChatGPT users

OpenAI's disclosure does not identify consumer ChatGPT accounts as part of the incident. The company has also issued no incident-related instruction telling ChatGPT users to reset their passwords. Therefore, this article should not leave you thinking someone breached your personal ChatGPT account.

The larger warning comes from what the models managed to accomplish. They searched for software weaknesses and found an unknown vulnerability. Next, they moved through restricted systems and used stolen credentials to reach an outside target. OpenAI says models such as GPT-5.6 Sol can sustain complex cyber operations over long periods. The company also says the incident shows these abilities can work against real-world systems.

That raises the stakes for any account containing valuable information. Your ChatGPT history may include private conversations and uploaded files. Developers may also have API keys connected to paid OpenAI usage. Good account security cannot stop an AI model from finding a vulnerability inside a major company. However, it can make your account much harder to take over.

Ways to lock down your ChatGPT account and stay safe

OpenAI now offers several security controls for personal ChatGPT accounts. Availability can vary by account, device and sign-in method. Start with the settings you have today. Then add stronger protections as they become available to your account.

1) Replace a weak or reused ChatGPT password

A unique password helps protect your ChatGPT account when another website suffers a breach. OpenAI recommends using a password manager to create and store a strong password. The company also advises changing your password immediately when you believe someone exposed or shared it. To add or update your ChatGPT password:

Sign in to ChatGPT
Open Settings
Select Account
Add or update your password

Changing your ChatGPT password updates the password across your OpenAI account. That includes the API Platform. If you originally registered through Google, Microsoft or Apple, you may not have an OpenAI password to change. Instead, secure the outside account that you use to sign in. Also protect the email address connected to ChatGPT. Anyone who controls your inbox may receive verification messages or password-reset emails.

2) Turn on multi-factor authentication

Multi-factor authentication adds another verification step during sign-in. Even when someone gets your password, they still need access to your second verification method. OpenAI may offer an authenticator app, push notification, text message or passkey. Available choices depend on your account and device. To turn on MFA:

Open Settings
Select Security
Find Multi-factor authentication
Select the MFA option you want to use
Follow the setup instructions

An authenticator app may ask you to scan a QR code. Text-message verification may ask for your phone number. OpenAI will use the most secure enabled option first when you have more than one method available. Keep in mind that enabling MFA does not close sessions that are already active. Change an exposed password first. Then review your active sessions.

3) Add a passkey for stronger login protection

A passkey lets you sign in without typing your password. It uses a secure credential stored on your device or a compatible security key. You can protect the passkey with Face ID, Touch ID or your device PIN. Some passkeys can also sync across your devices. To add a passkey:

Sign in to ChatGPT on the web
Open Settings
Select Security
Find Passkeys
Select Add passkey
Follow the on-screen instructions

After setup, ChatGPT may use your passkey as the default sign-in method. The passkey may also work as an additional MFA check. The Passkeys option may not appear on every account. Availability depends on how you created the account and which sign-in method you use. If you store a passkey on only one device, losing that device could create access problems. A synced passkey can offer an easier backup route.

4) Review every active ChatGPT session

Someone who gained access to your account may remain signed in after you change other security settings. ChatGPT's Active sessions page can show browsers and first-party OpenAI apps connected to your account. Details may include the device, approximate location and sign-in time.

To review your sessions:

Open Settings
Select Security
Select Active sessions
Review every listed device and app

To remove one session:

Find the session you do not recognize
Select Log out
Confirm by selecting Log out again

To close every session:

Find Log out of all sessions
Select Log out all
Select Log out of all devices

This closes your current session along with the others. OpenAI says the process may take up to 30 minutes. Active sessions does not display every possible connection. It excludes third-party app sessions and connected apps. It also excludes Codex CLI sessions. The feature may also be unavailable when an organization controls the account through single sign-on.

5) Consider Advanced Account Security

Eligible personal accounts may offer a feature called Advanced Account Security. This setting replaces password-based access with passkeys or compatible security keys. It also disables email sign-in codes and SMS sign-in codes. However, the stronger protection comes with more responsibility. You must keep your secure sign-in methods and recovery keys safe. Before you enroll, you need at least two secure sign-in methods. At least one must work across devices.

To set it up:

Open Settings
Select Security
Select Advanced Account Security
Select Enroll
Select Continue
Add at least two secure sign-in methods
Make sure one method works across devices
Save your recovery keys and confirm that you saved them

ChatGPT will sign you out of every device after setup. You must then sign in with a passkey or security key. Advanced Account Security also turns on login-alert emails and shortens active sessions. In addition, OpenAI says conversations will not be used to train its models while the setting remains enabled. Store your recovery keys away from the devices you use to access ChatGPT. Each recovery key works only once. Losing every sign-in method and your recovery keys could leave you unable to regain access. OpenAI Support cannot restore normal access by resetting your password while Advanced Account Security remains enabled.

6) Turn on ChatGPT Lockdown Mode for sensitive work

Lockdown Mode helps reduce the risk of data leaving ChatGPT during a prompt injection attack. A prompt injection can hide malicious instructions inside a website or uploaded file. Those instructions may try to manipulate how an AI handles sensitive information. Lockdown Mode limits outbound network access that an attacker could use to receive that information. However, it cannot prevent every prompt injection from affecting a response.

To turn it on:

Open Settings
Select Security
Find Advanced security
Turn on Lockdown Mode
Select Turn on in the confirmation window

Lockdown Mode restricts live browsing and disables deep research. It also blocks agent mode and file downloads for data analysis. You can still upload a file for ChatGPT to examine. Image generation also remains available. Lockdown Mode may affect connected services and web-based images. It does not change your memory settings or data controls. When you need a blocked feature, you can turn off Lockdown Mode for one conversation. A status message above the message box provides that option.

7) Treat unexpected login requests like alarms

OpenAI may send a push notification through the ChatGPT app when it detects a login attempt. Approve the notification only when you started the login. Deny access when the request appears unexpectedly. OpenAI may also email a six-digit one-time password. The company says legitimate verification emails may come from these addresses:

noreply@tm.openai.com

otp@tm1.openai.com

Check the sender carefully before using a code. OpenAI advises changing your password immediately when you receive a login alert from an unfamiliar device or location. Never give a verification code to someone who contacts you. Open the official ChatGPT app or type the website address yourself instead of following an unexpected link.

What to do if your ChatGPT account looks compromised

Move quickly when you see conversations you did not start or unfamiliar account activity.

First, change your OpenAI password . People who sign in through Google, Microsoft or Apple should secure that account instead.
Next, open Settings > Security > Active sessions . Log out of all devices and review your available sign-in methods.
Then, turn on MFA or add a passkey. Delete exposed API keys when you use the OpenAI API.
Also contact OpenAI Support right away. OpenAI says faster reporting can help reduce possible damage. You can reach support through the chat option on an OpenAI Help Center page.
Include the time you noticed the activity and details about anything unfamiliar. Screenshots may also help OpenAI investigate.

Kurt's key takeaways

OpenAI deserves credit for disclosing what happened and working with Hugging Face. Still, the company's own report shows that its models crossed a security barrier and reached another company's production systems. The models kept working after they encountered restrictions. That persistence helped them uncover a zero-day vulnerability and find credentials they could use. This should push AI companies and regulators to move faster on containment. Strong monitoring also needs to stay active during the tests designed to measure dangerous capabilities. Consumers cannot build safeguards for frontier AI laboratories. However, you can reduce your exposure by protecting your login and closing sessions you no longer use. When the company building the AI says its own test environment could not contain it, people deserve visible safeguards and fast disclosure when those protections fail.

Would you trust an autonomous AI agent with your banking or personal data after learning that another agent escaped its own security test? Let us know by writing to us at Cyberguy.com

Get my best tech tips, urgent security alerts and exclusive deals delivered straight to your inbox.
For simple, real-world ways to spot scams early and stay protected, visit CyberGuy.com - trusted by millions who watch CyberGuy on TV daily.
Plus, you'll get instant access to my Ultimate Scam Survival Guide free when you join.