I think the Three Laws is a good start but honestly, binding an AI's behavior to force it not to act unlawfully would close a lot of loopholes (or find legal loopholes, I suppose, which would be interesting.)