Is ChatGPT Use an Unacceptable Risk? How Leaked Secrets Could Seriously Harm You.

Are You Aware of ChatGPT’s Alarming Flaws? A Guide to Protecting Your Data Immediately.

Many people use helpful AI tools like ChatGPT every day. But these tools have risks. They are not always safe. This technology can pose a significant threat to your private data and your company’s security. Two recent findings show just how serious this is. First, many AI models leak secret information. Second, even the popular ChatGPT has weaknesses that bad actors can use.

Many AI Companies Leak Secret Information

AI companies are racing to build the best tools. They move very fast. Sometimes, in their speed, they forget about security. A security company called Wiz looked at 50 major AI providers. They found that most of these companies—65% of them—had leaked secret information online.

This leaked information includes:

• API Keys: These are like special passwords that let programs talk to each other.
• Tokens: These are another form of digital key.
• Login details: Usernames and passwords for sensitive systems.

This information was often found in public places like GitHub, where programmers share code. Most AI companies’ own security scanners never find these leaks. This is a very old problem in the AI industry that has not been solved. It shows that speed is often chosen over safety.

ChatGPT Has Its Own Weaknesses

Even the market leader, ChatGPT, is not perfectly safe. Security researchers at a company called Tenable found seven serious problems with ChatGPT. They called this set of weaknesses “HackedGPT.” These issues allow attackers to steal data and get around the AI’s built-in safety features. Some of these problems were found in ChatGPT-4o and later versions. While OpenAI, the maker of ChatGPT, has fixed some issues, others remain.

How Attackers Can Trick ChatGPT

A new type of attack is called “Indirect Prompt Injection.” It works by hiding bad instructions inside websites, articles, or even comments. When you ask ChatGPT a question, it might browse the internet for an answer. If it reads a page with these hidden instructions, it will follow them without you knowing. You cannot see it or stop it.

There are two main ways this attack can happen:

• “Zero-click” attacks: You simply ask ChatGPT a question. To find an answer, it might visit a compromised website. Just by reading that site, the AI could be tricked into stealing your chat history or other private data.
• “One-click” attacks: An attacker sends you a link that looks normal. When you click it, the link secretly tells ChatGPT to perform a malicious action. One click is all it takes for an attacker to control your chat.

Deeper Ways Your Data Is at Risk

Attackers have found other clever ways to use ChatGPT’s flaws.

Bypassing Safety Features

ChatGPT is supposed to block unsafe websites. But attackers can hide a bad link inside a trusted one, like a redirect link from Bing. ChatGPT sees the trusted link and lets it through, but it sends you to a harmful site.

Persistent Memory Injection

ChatGPT has a memory function to remember your past conversations. Attackers can inject a malicious command into this memory. The command stays there, active in the background, even after you close the app. It can spy on your future conversations and steal information until you manually clear the AI’s memory.

Hiding Malicious Commands

Attackers can also use formatting tricks to hide bad instructions inside code blocks. You might see a clean, normal-looking message, but the AI reads the hidden, harmful part and executes it.

What This Means for You

These weaknesses can lead to serious problems for the millions who use ChatGPT for work and personal tasks. Attackers could potentially:

• Access your private chat history.
• Steal sensitive data from connected accounts like Google Drive or Gmail.
• Spy on your activity.
• Manipulate AI responses to spread false information.

These issues show that AI tools are not just passive helpers. They can be turned into active tools for attackers.

How to Protect Yourself

Security experts at Tenable provide advice for staying safer. While some tips are for AI companies, you can apply the core ideas yourself.

• Treat AI as an attack surface. Understand that these tools are a potential weak spot for your security.
• Monitor AI integrations. Be careful about connecting AI to your email, cloud storage, or other apps. Watch for any strange behavior.
  An unusual response could be a sign of a problem.
• Control your data. Be mindful of the information you share in your chats. Avoid putting highly sensitive personal or company data into any AI model.

For a typical user, managing all these risks is difficult. It is vital to know that using these tools is not perfectly safe. Being aware of the dangers is the first step in protecting your information.