Lock down your ChatGPT account before the next AI attack

OpenAI has admitted that a operation of its precocious AI models recovered a mode retired of a locked-down trial situation and compromised different company's systems. That should marque each ChatGPT idiosyncratic halt and wage attention. The models were expected to enactment wrong a tightly controlled integer sandbox with nary unfastened net access. Instead, they recovered a weakness, got online and compromised systems belonging to Hugging Face, a large level for AI models and datasets.

OpenAI described the breach arsenic an "unprecedented cyber incident" involving "state-of-the-art cyber capabilities." The institution says the strategy included GPT-5.6 Sol and an adjacent much almighty exemplary inactive being tested down closed doors. Here is what concerns me. OpenAI built the models and controlled the investigating environment, yet the systems inactive recovered a mode astir the safeguards designed to incorporate them.

The AI models were trying to implicit a cybersecurity challenge. They seemingly had nary extremity of damaging Hugging Face. However, they kept pushing until they recovered vulnerabilities and crossed into different company's infrastructure. That should beryllium a wake-up call. You cannot power what happens wrong OpenAI's labs, but you tin instrumentality a fewer minutes to fastener down your ChatGPT relationship earlier an AI-powered menace comes looking for your password, idiosyncratic files oregon backstage conversations.

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

Our escaped CyberGuy Live class, "Sick of Spam?" , has ended, but you tin inactive ticker the afloat replay and download our spam-stopping checklist. Kurt "CyberGuy" Knutsson walks you measurement by measurement done elemental ways to trim robocalls, spam texts, junk email and unwanted messages. You’ll besides larn however to curb governmental texts, cleanable up your inbox and spot messages that could enactment your idiosyncratic accusation astatine risk.

Get the escaped replay and checklist present astatine CyberGuyLive.com.

OPENAI'S SAM ALTMAN WANTS TO NEGOTIATE A 5% STAKE IN COMPANY FOR US IF COMPETITORS AGREE TO KEY PROVISION

OpenAI says its precocious models escaped a restricted investigating situation and compromised Hugging Face portion pursuing an interior information benchmark. (Kurt "CyberGuy" Knutsson)

OpenAI says its models escaped during a cyber test

OpenAI designed the interior valuation to measurement however good its models could find and exploit hard information vulnerabilities. The institution ran the trial without immoderate of the accumulation information systems that usually artifact high-risk cyber activity. OpenAI said researchers removed those protections truthful they could measurement the models' maximum capabilities.

The models operated wrong a restricted investigating environment. Internet entree was constricted to an internally hosted work that acted arsenic a proxy for bundle packages. However, the models recovered a antecedently chartless vulnerability successful that service. They exploited the weakness and moved done OpenAI's probe situation until they reached a machine with net access.

Once online, the models identified Hugging Face arsenic a imaginable root of answers for the ExploitGym information benchmark. The models past searched for accusation that could assistance them implicit the test. OpenAI said they combined respective onslaught methods. Those methods included stolen credentials and antecedently chartless vulnerabilities.

In 1 case, the models recovered a way that allowed distant codification execution connected Hugging Face servers. That gave them the quality to tally codification connected different company's infrastructure. OpenAI said the models remained focused connected completing the evaluation. Yet their constrictive extremity inactive led them to transverse information boundaries and compromise an extracurricular company.

OpenAI says the incidental exposed a increasing spread betwixt what precocious models tin bash and the safeguards designed to incorporate them. "The superior acquisition from this incidental is that exemplary information and information indispensable support gait with rapidly advancing capabilities," OpenAI said successful its incidental report.

Hugging Face detected an autonomous AI intrusion

Hugging Face archetypal disclosed the breach connected July 16, 2026. The institution said an autonomous AI cause strategy carried retired the intrusion from opening to end. The onslaught progressive thousands of automated actions crossed short-lived integer environments.

Hugging Face identified unauthorized entree to a constricted acceptable of interior datasets. The attacker besides accessed respective credentials utilized by its services. However, the institution recovered nary grounds that anyone altered its nationalist models oregon user-facing datasets. It besides verified that its bundle proviso concatenation remained clean.

Hugging Face closed the vulnerabilities utilized for the archetypal access. It besides rebuilt affected systems and rotated exposed credentials. Meanwhile, the institution advised Hugging Face customers to rotate their entree tokens and reappraisal caller activity. That guidance applies to Hugging Face accounts alternatively than user ChatGPT accounts. OpenAI aboriginal determined that its models caused the enactment during its interior evaluation. The 2 companies proceed to analyse the incidental together.

CyberGuy reached retired to OpenAI and Hugging Face for remark but did not person a effect earlier publication.

FBI HELPS TAKE DOWN AI PHISHING RING

Passkeys and multi-factor authentication tin marque it harder for idiosyncratic to participate your ChatGPT relationship with a stolen password. (Kurt "CyberGuy" Knutsson)

What the AI sandbox flight means for ChatGPT users

OpenAI's disclosure does not place user ChatGPT accounts arsenic portion of the incident. The institution has besides issued nary incident-related acquisition telling ChatGPT users to reset their passwords. Therefore, this nonfiction should not permission you reasoning idiosyncratic breached your idiosyncratic ChatGPT account.

The larger informing comes from what the models managed to accomplish. They searched for bundle weaknesses and recovered an chartless vulnerability. Next, they moved done restricted systems and utilized stolen credentials to scope an extracurricular target. OpenAI says models specified arsenic GPT-5.6 Sol tin prolong analyzable cyber operations implicit agelong periods. The institution besides says the incidental shows these abilities tin enactment against real-world systems.

That raises the stakes for immoderate relationship containing invaluable information. Your ChatGPT past whitethorn see backstage conversations and uploaded files. Developers whitethorn besides person API keys connected to paid OpenAI usage. Good relationship information cannot halt an AI exemplary from uncovering a vulnerability wrong a large company. However, it tin marque your relationship overmuch harder to instrumentality over.

AMAZON RECALL TEXT SCAM COMES WITH RED FLAGS

Reviewing progressive ChatGPT sessions tin assistance you spot unfamiliar devices and region entree you bash not recognize. (Chris Ratcliffe/Bloomberg via Getty Images)

Ways to fastener down your ChatGPT relationship and enactment safe

OpenAI present offers respective information controls for idiosyncratic ChatGPT accounts. Availability tin alteration by account, instrumentality and sign-in method. Start with the settings you person today. Then adhd stronger protections arsenic they go disposable to your account.

1) Replace a anemic oregon reused ChatGPT password

A unsocial password helps support your ChatGPT relationship erstwhile different website suffers a breach. OpenAI recommends utilizing a password manager to make and store a beardown password. The institution besides advises changing your password instantly erstwhile you judge idiosyncratic exposed oregon shared it. To adhd oregon update your ChatGPT password:

• Sign successful to ChatGPT
• Open Settings
• Select Account
• Add oregon update your password

Changing your ChatGPT password updates the password crossed your OpenAI account. That includes the API Platform. If you primitively registered done Google, Microsoft oregon Apple, you whitethorn not person an OpenAI password to change. Instead, unafraid the extracurricular relationship that you usage to motion in. Also support the email code connected to ChatGPT. Anyone who controls your inbox whitethorn person verification messages oregon password-reset emails.

2) Turn connected multi-factor authentication

Multi-factor authentication adds different verification measurement during sign-in. Even erstwhile idiosyncratic gets your password, they inactive request entree to your 2nd verification method. OpenAI whitethorn connection an authenticator app, propulsion notification, substance connection oregon passkey. Available choices beryllium connected your relationship and device. To crook connected MFA:

• Open Settings
• Select Security
• Find Multi-factor authentication
• Select the MFA enactment you privation to use
• Follow the setup instructions

An authenticator app whitethorn inquire you to scan a QR code. Text-message verification whitethorn inquire for your telephone number. OpenAI volition usage the astir unafraid enabled enactment archetypal erstwhile you person much than 1 method available. Keep successful caput that enabling MFA does not adjacent sessions that are already active. Change an exposed password first. Then reappraisal your progressive sessions.

3) Add a passkey for stronger login protection

A passkey lets you motion successful without typing your password. It uses a unafraid credential stored connected your instrumentality oregon a compatible information key. You tin support the passkey with Face ID, Touch ID oregon your instrumentality PIN. Some passkeys tin besides sync crossed your devices. To adhd a passkey:

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

After setup, ChatGPT whitethorn usage your passkey arsenic the default sign-in method. The passkey whitethorn besides enactment arsenic an further MFA check. The Passkeys enactment whitethorn not look connected each account. Availability depends connected however you created the relationship and which sign-in method you use. If you store a passkey connected lone 1 device, losing that instrumentality could make entree problems. A synced passkey tin connection an easier backup route.

4) Review each progressive ChatGPT session

Someone who gained entree to your relationship whitethorn stay signed successful aft you alteration different information settings. ChatGPT's Active sessions leafage tin amusement browsers and first-party OpenAI apps connected to your account. Details whitethorn see the device, approximate determination and sign-in time.

To reappraisal your sessions:

• Open Settings
• Select Security
• Select Active sessions
• Review each listed instrumentality and app

To region 1 session:

• Find the league you bash not recognize
• Select Log out
• Confirm by selecting Log retired again

To adjacent each session:

• Find Log retired of each sessions
• Select Log retired all
• Select Log retired of each devices

This closes your existent league on with the others. OpenAI says the process whitethorn instrumentality up to 30 minutes. Active sessions does not show each imaginable connection. It excludes third-party app sessions and connected apps. It besides excludes Codex CLI sessions. The diagnostic whitethorn besides beryllium unavailable erstwhile an enactment controls the relationship done azygous sign-on.

5) Consider Advanced Account Security

Eligible idiosyncratic accounts whitethorn connection a diagnostic called Advanced Account Security. This mounting replaces password-based entree with passkeys oregon compatible information keys. It besides disables email sign-in codes and SMS sign-in codes. However, the stronger extortion comes with much responsibility. You indispensable support your unafraid sign-in methods and betterment keys safe. Before you enroll, you request astatine slightest 2 unafraid sign-in methods. At slightest 1 indispensable enactment crossed devices.

To acceptable it up:

• Open Settings
• Select Security
• Select Advanced Account Security
• Select Enroll
• Select Continue
• Add astatine slightest 2 unafraid sign-in methods
• Make definite 1 method works crossed devices
• Save your betterment keys and corroborate that you saved them

ChatGPT volition motion you retired of each instrumentality aft setup. You indispensable past motion successful with a passkey oregon information key. Advanced Account Security besides turns connected login-alert emails and shortens progressive sessions. In addition, OpenAI says conversations volition not beryllium utilized to bid its models portion the mounting remains enabled. Store your betterment keys distant from the devices you usage to entree ChatGPT. Each betterment cardinal works lone once. Losing each sign-in method and your betterment keys could permission you incapable to regain access. OpenAI Support cannot reconstruct mean entree by resetting your password portion Advanced Account Security remains enabled.

6) Turn connected ChatGPT Lockdown Mode for delicate work

Lockdown Mode helps trim the hazard of information leaving ChatGPT during a punctual injection attack. A punctual injection tin fell malicious instructions wrong a website oregon uploaded file. Those instructions whitethorn effort to manipulate however an AI handles delicate information. Lockdown Mode limits outbound web entree that an attacker could usage to person that information. However, it cannot forestall each punctual injection from affecting a response.

To crook it on:

• Open Settings
• Select Security
• Find Advanced security
• Turn connected Lockdown Mode
• Select Turn connected successful the confirmation window

Lockdown Mode restricts unrecorded browsing and disables heavy research. It besides blocks cause mode and record downloads for information analysis. You tin inactive upload a record for ChatGPT to examine. Image procreation besides remains available. Lockdown Mode whitethorn impact connected services and web-based images. It does not alteration your representation settings oregon information controls. When you request a blocked feature, you tin crook disconnected Lockdown Mode for 1 conversation. A presumption connection supra the connection container provides that option.

7) Treat unexpected login requests similar alarms

OpenAI whitethorn nonstop a propulsion notification done the ChatGPT app erstwhile it detects a login attempt. Approve the notification lone erstwhile you started the login. Deny entree erstwhile the petition appears unexpectedly. OpenAI whitethorn besides email a six-digit one-time password. The institution says morganatic verification emails whitethorn travel from these addresses:

[email protected]

[email protected]

Check the sender cautiously earlier utilizing a code. OpenAI advises changing your password instantly erstwhile you person a login alert from an unfamiliar instrumentality oregon location. Never springiness a verification codification to idiosyncratic who contacts you. Open the authoritative ChatGPT app oregon benignant the website code yourself alternatively of pursuing an unexpected link.

What to bash if your ChatGPT relationship looks compromised

Move rapidly erstwhile you spot conversations you did not commencement oregon unfamiliar relationship activity.

• First, alteration your OpenAI password . People who motion successful done Google, Microsoft oregon Apple should unafraid that relationship instead.
• Next, unfastened Settings > Security > Active sessions . Log retired of each devices and reappraisal your disposable sign-in methods.
• Then, crook connected MFA oregon adhd a passkey. Delete exposed API keys erstwhile you usage the OpenAI API.
• Also interaction OpenAI Support close away. OpenAI says faster reporting tin assistance trim imaginable damage. You tin scope enactment done the chat enactment connected an OpenAI Help Center page.
• Include the clip you noticed the enactment and details astir thing unfamiliar. Screenshots whitethorn besides assistance OpenAI investigate.

Kurt's cardinal takeaways

OpenAI deserves recognition for disclosing what happened and moving with Hugging Face. Still, the company's ain study shows that its models crossed a information obstruction and reached different company's accumulation systems. The models kept moving aft they encountered restrictions. That persistence helped them uncover a zero-day vulnerability and find credentials they could use. This should propulsion AI companies and regulators to determination faster connected containment. Strong monitoring besides needs to enactment progressive during the tests designed to measurement unsafe capabilities. Consumers cannot physique safeguards for frontier AI laboratories. However, you tin trim your vulnerability by protecting your login and closing sessions you nary longer use. When the institution gathering the AI says its ain trial situation could not incorporate it, radical merit disposable safeguards and accelerated disclosure erstwhile those protections fail.

Would you spot an autonomous AI cause with your banking oregon idiosyncratic information aft learning that different cause escaped its ain information test? Let america cognize by penning to america astatine Cyberguy.com

CLICK HERE TO DOWNLOAD THE FOX NEWS APP

Sign up for my FREE CyberGuy Report

• Get my champion tech tips, urgent information alerts and exclusive deals delivered consecutive to your inbox.
• For simple, real-world ways to spot scams aboriginal and enactment protected, sojourn CyberGuy.com - trusted by millions who ticker CyberGuy connected TV daily.
• Plus, you'll get instant entree to my Ultimate Scam Survival Guide escaped erstwhile you join.

Copyright 2026 CyberGuy.com. All rights reserved.