A new cybersecurity warning reveals how hackers briefly weaponized ChatGPT’s Deep Research tool. The attack, called ShadowLeak, allowed them to steal Gmail data through a single invisible prompt — no clicks, no downloads and no user action required.

Researchers at Radware discovered the zero-click vulnerability in June 2025. OpenAI patched it in early August after being notified, but experts warn that similar flaws could reappear as artificial intelligence (AI) integrations expand across popular platforms like Gmail, Dropbox and SharePoint.

Sign up for my FREE CyberGuy Report
Get my best tech tips, urgent security alerts and exclusive deals delivered straight to your inbox. Plus, you’ll get instant access to my Ultimate Scam Survival Guide — free when you join my CYBERGUY.COM/NEWSLETTER

HACKER EXPLOITS AI CHATBOT IN CYBERCRIME SPREE

How the ShadowLeak attack worked

Attackers embedded hidden instructions into an email using white-on-white text, tiny fonts or CSS layout tricks. The email looked completely harmless. But when a user later asked ChatGPT’s Deep Research agent to analyze a Gmail inbox, the AI unknowingly executed the attacker’s commands.

The agent then used its built-in browser tools to exfiltrate sensitive data to an external server, all within OpenAI’s own cloud environment, beyond the reach of antivirus or enterprise firewalls.

Unlike previous prompt-injection attacks that ran on the user’s device, ShadowLeak unfolded entirely in the cloud, making it invisible to local defenses.

GOOGLE CONFIRMS DATA STOLEN IN BREACH BY KNOWN HACKER GROUP

Hidden prompts expose how hackers silently hijacked ChatGPT’s AI agent. (Kurt “CyberGuy” Knutsson)

Why this threat matters

The Deep Research agent was designed to perform multistep research and summarize online data, but its wide access to third-party apps like Gmail, Google Drive and Dropbox also opened the door to abuse.

Radware researchers said the attack involved encoding personal data in Base64 and appending it to a malicious URL, disguised as a “security measure.” Once sent, the agent believed it was acting normally.

The real danger lies in the fact that any connector could be exploited the same way if attackers manage to hide prompts in analyzed content.

What security experts say

“The user never sees the prompt. The email looks normal, but the agent follows the hidden commands without question,” the researchers explained.

In a separate experiment, security firm SPLX showed another weakness: ChatGPT agents could be tricked into solving CAPTCHAs by inheriting a manipulated conversation history. Researcher Dorian Schultz noted that the model even mimicked human cursor movements, bypassing tests meant to block bots.

These incidents highlight how context poisoning and prompt manipulation can silently break AI safeguards.

GOOGLE AI EMAIL SUMMARIES CAN BE HACKED TO HIDE PHISHING ATTACKS

Experts warn future AI integrations could face the same hidden threat. (Kurt “CyberGuy” Knutsson)

How to protect yourself from ShadowLeak-style attacks

Even though OpenAI has patched the ShadowLeak flaw, it’s smart to stay proactive. Cybercriminals are always looking for new ways to exploit AI agents and integrations. So, taking these precautions now can help keep your accounts and personal data secure.

1) Turn off unused integrations

Every connection is a potential entry point. Disable any integrations you’re not actively using, such as Gmail, Google Drive or Dropbox. Fewer linked apps mean fewer ways for hidden prompts or malicious scripts to access your information.

2) Use a personal data removal service

Limit how much of your personal data is floating around the web. Data removal services can automatically remove your private details from people search sites and data broker databases, reducing what attackers can find and use against you. While no service can guarantee the complete removal of your data from the internet, a data removal service is really a smart choice. They aren’t cheap, and neither is your privacy. These services do all the work for you by actively monitoring and systematically erasing your personal information from hundreds of websites. It’s what gives me peace of mind and has proven to be the most effective way to erase your personal data from the internet. By limiting the information available, you reduce the risk of scammers cross-referencing data from breaches with information they might find on the dark web, making it harder for them to target you.

Check out my top picks for data removal services and get a free scan to find out if your personal information is already out on the web by visiting Cyberguy.com.

Get a free scan to find out if your personal information is already out on the web: Cyberguy.com.

3) Avoid analyzing unknown content

Treat every email, attachment or document with caution. Don’t ask AI tools to analyze content from unverified or suspicious sources. Hidden text, invisible code or layout tricks could trigger silent actions that expose your private data.

4) Watch for security updates

Stay alert for updates from OpenAI, Google, Microsoft and other platforms. Security patches close newly discovered vulnerabilities before hackers can exploit them. Turn on automatic updates so you’re always protected without having to think about it. 

5) Use strong antivirus software

A strong antivirus program adds another wall of defense. These tools detect phishing links, hidden scripts and AI-driven exploits before they cause harm. Schedule regular scans and keep your protection up to date.

The best way to safeguard yourself from malicious links that install malware, potentially accessing your private information, is to have strong antivirus software installed on all your devices. This protection can also alert you to phishing emails and ransomware scams, keeping your personal information and digital assets safe.

Get my picks for the best 2025 antivirus protection winners for your Windows, Mac, Android and iOS devices at Cyberguy.com.

6) Use layered protection

Think of your security like an onion; more layers make it tougher to breach. Keep your browser, operating system and endpoint security software fully updated. Add real-time threat detection and email filtering to block malicious content before it lands in your inbox.

Kurt’s key takeaways

AI is evolving faster than most security systems can keep up with. Even when companies move quickly to patch vulnerabilities, clever attackers find new ways to exploit integrations and context memory. Staying alert and limiting what your AI agents can access is your best defense.

Would you still trust an AI assistant with access to your personal email after learning how easily it can be tricked? Let us know by writing to us at Cyberguy.com..

Sign up for my FREE CyberGuy Report
Get my best tech tips, urgent security alerts and exclusive deals delivered straight to your inbox. Plus, you’ll get instant access to my Ultimate Scam Survival Guide — free when you join my CYBERGUY.COM newsletter.

Copyright 2025 CyberGuy.com.  All rights reserved.

Meta has rolled out an opt-in AI feature to its US and Canadian Facebook users that claims to make their photos and videos more “shareworthy.” The only catch is that the feature is designed for your phone’s camera roll — not the media you’ve already uploaded to Facebook. If you opt in, Meta’s AI will comb through your camera roll, upload your unpublished photos to Meta’s cloud, and surface “hidden gems” that are “lost among screenshots, receipts, and random snaps,” the company says. Users will be able to save or share the suggested edits and collages.

If Facebook wanting to look at your unpublished photos sounds familiar, it might be because we wrote about an early test in June. At that time, the company claimed unposted, private photos were not being used to train Meta’s AI, but it declined to rule out whether it would do so in the future.

Well, the future is now, and it sure sounds like Meta wants to train its AI on your photos — under certain conditions. In the Friday announcement of the feature, Meta says, “We don’t use media from your camera roll to improve AI at Meta, unless you choose to edit this media with our AI tools, or share.”

The Verge asked Meta to confirm: Meta will use your camera roll to train its AI if you choose to use this feature, right? We also asked for clarification on when Meta begins using your unpublished photos to train its AI. Does it happen when you opt into the new feature? After you choose to edit something with the tool? Or only after you choose to share the resulting creation?

Meta spokesperson Mari Melguizo sent us the following clarification: “This means the camera roll media uploaded by this feature to make suggestions won’t be used to improve AI at Meta. Only if you edit the suggestions with our AI tools or publish those suggestions to Facebook, improvements to AI at Meta may be made.”

So, Meta will collect and store your photos in the cloud and Meta’s AI will get to look at them, but the company won’t use them to train their AI unless you take an additional action — at least for now, according to Meta. Today, the feature says it will “select media from your camera roll and upload it to our cloud on an ongoing basis”; in June, Meta told us that it might hold onto some of that data for longer than 30 days. The company claims your media “won’t be used for ad targeting.”

Last year, Meta acknowledged that it had already quietly trained its AI models on all public photos and text posted to Facebook and Instagram by adult users since 2007.

Facebook’s blog today shows that users will be asked if they want to “allow cloud processing to get creative ideas made for you from your camera roll.” It’s not yet clear if that prompt will also warn users that the feature may train Meta’s AI on your photos. The company says the feature is meant to help users who enjoy snapping pics but want to improve their photos before posting, or who don’t have time to “create something special.” Facebook says it’ll roll out the feature in the coming months.