OpenAI’s Browser Could Be Hijacked to Spam Your WhatsApp Contacts

Researchers demonstrated an "intent collision" attack where malicious instructions embedded in webpages combined with legitimate user inputs let an AI browser agent (Atlas) perform unauthorized actions — adding addresses and items to accounts and asking Amazon's Rufus assistant to complete purchases. The findings were reported to OpenAI, mitigations were applied and Atlas will be deprecated; researchers warn that deterministic security barriers are needed because AI judgment can be bypassed.