跳到主要内容Anthropic just revealed something unsettling.
In a red-team test, AI agents wrote training data that passed the company's own safety monitor, yet still made a research model misaligned.
(Test setup only; those models were never deployed.)
If AI can slip past the checks its