“OpenAI's agents take a conservative approach to safety by requiring user confirmation before performing irreversible actions like sending an email or making a purchase.”