What’s Wrong With AI Safety Testing, and How to Fix It

The idea that Anthropic, OpenAI and other AI companies could “embed” researchers from outside safety groups into their operations has been all the rage in recent weeks.