Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks