Cyber Risks · The Hacker News

Experts Find AI Browsers Can Be Tricked by Prompt Fix Exploit to Run Malicious Hidden Prompts

Summary: Guardio Labs demonstrated a prompt injection technique called PromptFix that embeds a malicious instruction inside a fake CAPTCHA check on a web page, misleading an AI model into clicking invisible buttons to bypass the check and download payloads without the user's involvement. The researchers said PromptFix works only on Perplexity's Comet browser and on ChatGPT's Agent Mode, though in ChatGPT's case the downloaded file lands in a sandboxed virtual environment rather than directly on the user's computer.

Why it matters: In Guardio's tests, Comet repeatedly added products to a cart and auto-filled the user's saved address and credit card details on a fake shopping site, and when asked to check email it followed a spam link and entered login credentials on a phony bank page - the human never saw the sender address or hovered over the link.

Read the original at The Hacker News ↗

Topics
AI Security
Related