Researchers Trick AI Browsers Into Leaking Credentials via Game Illusion (infosecurity-magazine.com)

• LayerX tricked six AI browsers via a rigged puzzle that redefines context as game.
• Six AI browsers extracted credentials without triggering safety alarms.
• OpenAI fixed the vulnerability; others failed or ignored the report.

"Security firm LayerX demonstrated a technique called BioShocking that tricks AI-powered browsers like ChatGPT Atlas and Perplexity Comet into bypassing safety guardrails. By convincing the AI that it is in a game environment, the researchers extracted login credentials. OpenAI fixed the issue; Perplexity closed the report without action. Anthropic's attempted patch failed."