Why we need a global framework with inbuilt safety mechanisms for AI