We Vetted 2000 AI Skills Before They Reached Developers — Lucas Palma, Nubank
Lucas Palma from Nubank describes how the bank built a tool called Skill Vector to control AI features — small code programs that extend what AI models can do — before they reach developers.
Each feature is first scanned with automated controls that search for dangerous commands, then reviewed by an AI model to catch problems that automation misses. After checking over 2000 features this way, they found real security issues that were fed into the bank's vulnerability management system. Nubank combined automated scanning with AI review, but needed to improve guidance and prevent developers from running uncontrolled features locally before approval.
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
