“The UK AI Security Institute conducted a long-term red teaming collaboration with Anthropic and OpenAI that found significantly more jailbreaks than a normal pre-deployment evaluation would have.”