The Anthropic–OpenAI feud and their Pentagon dispute expose a deeper problem with AI safety