The hidden risks of AI's increasingly helpful agents