Nubank's Playbook for AI Supply Chain Security That Most Teams Lack

Nubank’s product security manager reveals a systemic gap: AI skills—configuration files that guide agent behavior—behave like supply chain dependencies but lack traditional security vetting. After scanning 2,000 skills internally, they found over 1,500 risks, including credential leaks, destructive shell commands, and excessive permissions. The scale is non-trivial.

Teams likely treat AI skills as harmless plugins. That assumption is dangerous. Nubank built Skill Vector, a hybrid CI scanner using deterministic regex and LLM reviews, to catch unsafe instructions before skills reach the internal marketplace. The system blocked critical risks and forced remediation on hundreds of others.

The tradeoff is governance velocity. Local iteration is allowed, but CI gating creates friction. Developers want speed; security demands safety. Nubank’s solution—scanning locally, in PRs, and post-upload—balances both, but only for a single marketplace. External skill sources remain unvetted.

Engineering leaders should watch for this pattern. Skills, MCP servers, and agent rules are the new attack surface. Most organizations lack equivalent review pipelines. The lesson is clear: you must inspect the configuration layer, not just the generated code, or risk silent supply chain contamination.