The testing found the AI was capable of "extreme actions" if it thought its "self-preservation" was threatened.