AI Sidebar Spoofing Puts ChatGPT Atlas, Perplexity Comet and Other Browsers at Risk

SquareX researchers have demonstrated how malicious browser extensions can impersonate AI sidebar interfaces to carry out phishing and other attacks, in a method they call AI Sidebar Spoofing. The demonstrations targeted Perplexity’s Comet and ChatGPT Atlas, OpenAI’s web browser, but SquareX argues the flaw is systemic and that Edge, Brave and Firefox are also vulnerable.

The attack works by the malicious extension injecting JavaScript to create a fake, visually indistinguishable AI sidebar; when a user enters a prompt, the spoofed sidebar feeds the prompt to the real model but, for certain instructions, manipulates the response to include malicious steps. This can steer users toward phishing sites or instruct them to perform actions such as executing commands that enable malware deployment, including a reverse shell.

OpenAI has noted safeguards in Atlas, though the article cautions these protections are limited if social engineering drives users to install the extension or trust the fake sidebar, with the issues requiring user interaction to succeed. according to OpenAI.