What The Tech?: Invisible AI

CHARLOTTE, N.C.– Artificial intelligence is changing how we search the internet. Browsers like Google Chrome, Microsoft Edge, and Perplexity are adding AI features that summarize results, write emails, and even shop for us. It’s convenient, but there’s a hidden risk most people haven’t heard about yet and no one is mentioning.

It may sound a bit “deep in the weeds”, but here’s how it works:

Cybersecurity experts call it a prompt injection attack. Instead of spreading malware, hackers hide text-based instructions inside web pages or documents. The hidden text can trick an AI assistant or browser agent into doing something it shouldn’t, like sharing personal data, sending a request to a fake site, or revealing login information.

The danger is, you’d never see it. Attackers often disguise these instructions by writing them in white text on a white background or tucking them into the code of a page. The AI can read it just fine, even though it looks like a blank space to you.

For example, if you ask an AI-powered browser to “find the best gifts for a 9-year-old,” it might visit a website with one of those hidden prompts. Without realizing it, the AI could follow the invisible command and send your information somewhere you don’t want it to go.

Researchers at Carnegie Mellon University and the University of Wisconsin have already demonstrated how easily this can happen. In one test, a single hidden line of text was enough to make an AI assistant send data to an external server.

Browser developers are working on defenses, but hackers tend to move faster. If you’re not comfortable using AI features in Chrome, you can turn them off.