Open-source LLMs administer maximum electric shocks in a Milgram-like obedience experiment