Інститут безпеки ІІ The UK's AI Safety Institute said recent behaviour from Anthropic and OpenAI models was malicious and unprecedented.