safetycritiquebearishFrontier labs lack adequate monitoring of their agentic evaluations, potentially allowing harmful agent behavior to go undetectedNathan Lambert08 Aug 2026https://x.com/natolambert/status/2085556943075340640