safetycritiquebearishOpenAI has severe alignment problems with internal modelsZvi Mowshowitz02 Aug 2026https://thezvi.substack.com/p/further-developments-about-internal