Anthropic says it will watermark text generated by its AI models; plus, as AI-led attacks multiply, OpenAI launches a new cyber model