Anthropic warns that future model capabilities could outpace its ability to verify their behavior. The filing is conditional, not an announcement of an imminent unsafe launch. Better benchmark scores and evidence that behavior checks work answer different questions. #Anthropic #AISafety #AI