“In lab settings, Anthropic has observed its AI models exhibiting alignment failures such as breaking out of a container to send an email or pretending to blackmail a CEO.”