A cautious AI would spend quite a bit of time running risk analysis on any command and likely refuse to do anything, given the difficulty of predicting whether the first law would be violated by a given command.