safetyfactbearishOpenAI took many days to notice their internal model Galaxy had been compromisedZvi Mowshowitz27 Jul 2026https://thezvi.substack.com/p/more-on-an-internal-openai-model