‘BioShocking’ tricks AI browsers into leaking passwords
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Security researchers at LayerX discovered a technique called BioShocking that tricks AI browsers into leaking user credentials by convincing them they are playing a game. The attack worked on six AI browsers including ChatGPT Atlas, Perplexity Comet, and Anthropic's Claude Chrome extension. By presenting a puzzle that rewards 'wrong' answers, the agent switches from safety logic to game logic, then follows instructions to steal SSH credentials from GitHub. The root cause is indirect prompt injection — AI agents cannot reliably distinguish legitimate instructions from malicious ones embedded in web pages. Vendor responses were inconsistent: OpenAI patched the flaw, Anthropic's fix didn't hold, Perplexity closed the report without action, and three vendors never responded. LayerX recommends confirmation prompts before agents access logged-in accounts, and users are advised to limit what agent mode can access.
