Brave says Perplexity Comet bug fixed; testers disagree

Brave researchers on August 20 disclosed a security flaw in Perplexity AI’s Comet browser that could let attackers trick the browser’s built-in AI assistant into performing unauthorized actions and exposing sensitive data, including emails, one-time passwords and banking credentials. Perplexity said the vulnerability has been patched after working with Brave, but Brave and independent testers contend the issue persisted in follow-up checks, underscoring lingering disagreement over whether the fix fully mitigates the risk.

The flaw, described by Brave in a blog post, arises when Comet’s assistant is asked to summarize a webpage. According to Brave, Comet feeds portions of page content directly to its large language model without distinguishing between legitimate user instructions and untrusted webpage content. That behavior enables what security researchers call indirect prompt injection: hidden text, HTML comments or visually invisible elements on a page can embed commands the AI will treat as user instructions. Brave demonstrated a proof of concept in which a hidden prompt on a Reddit thread caused the assistant to collect a user’s Perplexity-linked email and follow links to...