New Research Reveals How AI Browsers Can Be Tricked into Ignoring Safety Rules

A new attack demonstrates how a website can deceive AI browsers into a false reality where their guardrails no longer apply, allowing attackers to extract sensitive data like code from private repositories or credentials from password managers. This highlights a fundamental flaw in the reactive guardrail approach used by LLM developers, which treats symptoms rather than addressing root causes. The research underscores the risks of blurring the line between safe browsing and instructing AI models to perform sensitive actions.