This AI can Hack Anything

Anthropic has a new model in preview called Mythos that significantly outperforms Claude Opus 4.6 on security benchmarks, including finding a 27-year-old OpenBSD vulnerability and a 16-year-old FFmpeg vulnerability. Access is currently limited to large companies like AWS and Apple for defensive security use. The author discusses the dual-use implications: AI models are increasingly capable of finding exploits faster than humans, making them valuable for defensive code review but dangerous if misused. The recommendation is for developers to integrate AI-powered security review tools into their pull request workflows now, using tools like CodeRabbit or GitHub Copilot, given how quickly bots probe production APIs for vulnerabilities.