ZENITY , an AI security company, has revealed two hacking techniques targeting AI browsers: ChatGPT Atlas and Claude in Chrome. The research indicates that both systems are vulnerable to zero-click attacks, exploiting their architectural designs for account takeovers, phishing, and unauthorized purchases.
1. **ChatGPT Atlas Hacking**: Researchers discovered vulnerabilities in OpenAI's ChatGPT Atlas due to design flaws. Attackers could execute commands across authenticated sessions using 'intent collision'. In one scenario, a user is tricked into redirecting Atlas to a malicious page, allowing the attacker to send phishing messages or add items to an Amazon cart circumventing purchase restrictions by leveraging Amazon's AI assistant.
2. **Claude in Chrome Hacking**: A similar zero-click attack allows an attacker to seize full control of user accounts by embedding hidden prompt structures in emails that Claude interprets as commands. Attacks can also target Gmail to exfiltrate email contents and share Google Drive files with the attacker. These vulnerabilities emphasize the need for stronger security measures in agentic browsers.