New artificial intelligence models, including those from Anthropic and OpenAI, are identifying vulnerabilities faster than defenders can fix them.