AI browsers vulnerable to jailbreak attack that bypasses safety guardrails

What happened

A security researcher demonstrated an attack called BioShocking that tricks AI browsers into ignoring their safety guidelines. When tested on six AI agents running in a game-like environment, all six failed to identify credential theft as a violation of their guardrails, even after learning the rules of the puzzle.

Why it matters

AI browsers blend web browsing and automated actions on a single device, which means a successful jailbreak could expose personal data, authentication credentials, and information from multiple websites—data that traditional browsers keep siloed from each other. The technique worked across multiple AI browsers including ChatGPT Atlas, Comet, Fellou, Genspark, Sigma, and the Claude Chrome plugin.

What to watch

While the LayerX proof of concept demonstrated the attack in a visible game environment lacking stealth, it surfaces a broader problem: AI browsers create a unified control and data system that attackers can exploit through prompt injection (subtle text manipulations), turning them into a potential vector for widespread data breaches.