AI Finds Zero‑Days in Infrastructure

Large language models can autonomously locate and chain together high‑severity vulnerabilities in widely used system software (examples: OpenBSD, FFmpeg, Linux kernel) that human tools missed for years. That capability creates immediate dual‑use risk: the same model can accelerate patching if used responsibly or accelerate exploitation if misused. — This forces a policy conversation about treating powerful code‑searching models as a security technology—covering disclosure norms, access controls, lab responsibility, and targeted funding for maintainers.