“Anthropic's model Mythos identified 250 security vulnerabilities in an open-source codebase where a prior model had only found 22.”