Anthropic recently discovered an important new fact about rogue AI agents: They, too, can be psychologically destroyed by CAPTCHA. During a security evaluation, Anthropic gave its Mythos 5 model a hacking task that was supposed to occur inside a sandbox....
