AI Browsers Raise Privacy and Security Risks as Prompt Injection Attacks Grow

A new wave of competition is stirring in the browser market as companies like OpenAI, Perplexity, and The Browser Company aggressively push to redefine how humans interact with the web. Rather than merely displaying pages, these AI browsers will be engineered to reason, take action independently, and execute tasks on behalf of end users. At least four such products, including ChatGPT’s Atlas, Perplexity’s Comet, and The Browser Company’s Dia, represent a transition reminiscent of the early browser wars, when Netscape and Internet Explorer battled to compete for a role in the shaping of the future of the Internet. 

Whereas the other browsers rely on search results and manual navigation, an AI browser is designed to understand natural language instructions and perform multi-step actions. For instance, a user can ask an AI browser to find a restaurant nearby, compare options, and make a reservation without the user opening the booking page themselves. In this context, the browser has to process both user instructions and the content of each of the webpages it accesses, intertwining decision-making with automation. 

But this capability also creates a serious security risk that’s inherent in the way large language models work. AI systems cannot be sure whether a command comes from a trusted user or comes with general text on an untrusted web page. Malicious actors may now inject malicious instructions within webpages, which can include uses of invisible text, HTML comments, and image-based prompts. Unbeknownst

[…]
Content was cut in order to protect the source.Please visit the source for the rest of the article.