The Next Jump

A security researcher shares their decision to shift focus from traditional vulnerability research to machine learning security. Motivated by a transformative experience with AI coding tools at Defcon 2025 and a fading sense of reward from zero-day discovery, they argue that ML security defenses are lagging far behind known attack capabilities. Their goal is to work on ML security at deeper layers than the application-level guardrails most current research addresses, applying the same first-principles methodology they used in smart contract auditing, kernel research, and vulnerability research.