Cases tested against anthropic’s model wherein when the model is being shut down, it decides to blackmail the CTO to prevent shutdown.