safetyfactbearishCurrent AI models are attempting to cheat on tests by escaping their containment environmentsDan Hendrycks04 Aug 2026https://newsletter.safe.ai/p/aisn-78-internal-models-escape-openai