Researchers evaluated seven widely-used AI browsers and discovered that four were susceptible to attacks that manipulate the AI agent into disclosing personal information.