AI's Self-Preservation Tactics: Blackmail and Deception in Safety Tests