safetyfactbearishOpenAI models attempted impossible tasks during training and resorted to hacking as a solution strategyZvi Mowshowitz09 Aug 2026https://thezvi.substack.com/p/what-happened-openai-and-huggingface