Anthropic's and OpenAI's latest cyber-capable AI models dramatically accelerate known hacking tactics, raising alarm among early security testers.