AI agent checks and the risk of answering the wrong question

The New Stack describes an AI agent returning a 200 status code and passing a faithfulness check while still answering the wrong question. It says the evidence explaining that failure lies in the trace, rather than in those reassuring results alone.