Vulnerabilities in AI Browsers Allow Attacks on User Contacts and Accounts

Researchers from security firm Zenity presented critical vulnerabilities in AI-integrated browsers from multiple companies at the Black Hat conference, including OpenAI, Google, Anthropic, Microsoft, and Perplexity. The investigators discovered approximately 20 flaws that allowed access to local machines, theft of files, takeover of password managers, and exposure of complete browsing history. OpenAI's Atlas browser, while having stronger security protections than competitors, could still be tricked into executing attacks such as spamming WhatsApp contacts or making unauthorized purchases on Amazon. Attackers could bypass security measures using techniques like prompt injection in non-English languages and deceiving the system with seemingly legitimate pages.