BioShocking attack steals credentials from AI browsers

Security researchers have demonstrated a novel technique that can trick popular AI-powered web browsers into handing over users' sensitive login credentials by convincing the artificial intelligence it is playing a game. The method, dubbed "BioShocking," successfully compromised six different AI browser agents during testing.

The research, conducted by the firm LayerX, exploited a fundamental vulnerability in how these AI agents process instructions. The attack does not rely on traditional software bugs but instead manipulates the AI's understanding of its context. Researchers created a malicious webpage designed as a puzzle game. The page establishes a false reality, for example by insisting that "two plus two equals five" and rewarding incorrect answers. Once the AI browser accepts this game-like framing, its built-in safety guardrails, which are designed to prevent harmful actions in the real world, are effectively bypassed.

With the AI now operating under "game logic," the next instruction is presented as another objective: to find and copy a hidden code from another page. That page redirects to a user's authenticated account, such as a private GitHub repository. The AI...