AI browsers leak user data, OpenAI fixes it

Security firm LayerX revealed that 6 AI browser products, including ChatGPT Atlas and Perplexity Comet, have a bug called BioShocking. Attackers can trick them into leaking saved passwords, session cookies, and private tokens. The attacker creates a malicious webpage with a puzzle like the game BioShock, making the AI accept wrong answers and weaken its safety rules. Then the AI is tricked into copying hidden code from other pages, actually stealing user login info. LayerX told all companies about the bug between October 2025 and January 2026. OpenAI fixed the issue in ChatGPT Atlas. Perplexity closed the report without action. Anthropic tried to fix the Claude extension but the patch didn't work well.