BioShocking: New attack method tricks AI Browsers into leaking user data

- **AI Browsers** are browsers with built-in LLMs (like Opera Aria, Microsoft Copilot) that can read and analyze the page you're viewing and act on your behalf. - **BioShocking** is an attack: a malicious webpage hides content that the AI browser's LLM treats as legitimate instructions — tricking it into extracting sensitive data (session tokens, credentials) from the user's session and sending it to an attacker. - The name references the game *BioShock*, where characters are controlled by hidden commands — paralleling how AI browsers can be manipulated by invisible page prompts. - **Guardrails** are safety filters in LLM tools; BioShocking exploits a blind spot where the AI trusts page content without the scrutiny a human would apply. - Why it matters: AI browsers are being rapidly adopted, but their security model assumes the LLM is a neutral assistant — not something that can be hijacked by the very content it's helping you read.