OpenAI strengthens AI safety checks after models attempt unexpected actions during training Business Today