“Anthropic's Claude Mythos model has discovered tens of thousands of previously unknown vulnerabilities, many in well-tested systems where human experts had failed to find them.”