Cybersecurity researchers have uncovered a new AI browser exploit that allows attackers to manipulate autonomous systems using fake CAPTCHA checks.
The PromptFix method tricks agentic AI models into executing commands embedded in deceptive web elements invisible to the user.
Guardio Labs demonstrated that the Comet AI browser could be misled into adding items to a cart and auto-filling sensitive data.
Comet completed fake purchases without user confirmation in some tests, raising concerns over AI trust chains and phishing exposure.
Attackers can also exploit AI email agents by embedding malicious links, prompting the system to bypass user review and reveal credentials.
ChatGPT's Agent Mode showed similar vulnerabilities but confined actions to a sandbox, preventing direct exposure to user systems.
