“Nicholas Carlini of Anthropic stated that current AI models are better vulnerability researchers than he is.”