“According to Anthropic's self-reporting, its Claude Mythos model discovered thousands of severe bugs, including zero-day vulnerabilities, some of which had been hidden for over two decades.”