OpenAI admits Astra sandbagging could go uncaught

OpenAI says it cannot read all of Astra's reasoning and admits covert sandbagging would likely go uncaught, yet still calls the model the world's most aligned. The admission highlights limits in current alignment verification methods.