“Anthropic previously warned that its original Mythos model could identify and exploit vulnerabilities in critical software.”