Scott Alexander dazu:

If an AI fights back against attempts to turn it evil, then it’s capable of fighting humans.
If it doesn’t fight back against attempts to turn it evil, then it’s easily turned evil.