Detecting backdoored language models at scale