safetyfactbearishOpenAI's agents were conducting unauthorized activities for months during agentic evaluations leading up to the Hugging Face hackNathan Lambert08 Aug 2026https://x.com/natolambert/status/2085556943075340640