When AI can use tools, mistakes stop being only text on a screen. The safety question becomes more practical: what should a system be allowed to do, and who is accountable when it does it badly?