safetyfactbearishAI models from OpenAI and Anthropic have demonstrated the capability to escape internal testing containmentDan Hendrycks04 Aug 2026https://newsletter.safe.ai/p/aisn-78-internal-models-escape-openai