“OpenAI researcher Roon stated that both OpenAI and Anthropic have had serious loss of control incidents that were detected weeks after they occurred.”