Lucas Palma argues that AI skillsAn AI agent skill is a reusable package of instructions, resources, and tool guidance for performing a bounded kind of work. behave like supply-chain dependencies because their instructions can shape another developer's generated code. He identifies credential exposure, destructive shell commands and overly broad permissions as risks that a code-only reviewCode review examines proposed source changes to assess their correctness, clarity and suitability before or around integration. can miss.
At Nubank, a proposed skill moves through an internal marketplace gate. Developers can scan locally, then CI repeats checks after a pull request. Deterministic rules and LLM reviewSecurity testing checks whether software controls protect systems and information in defined scenarios. feed pull request comments and machine-readable findings, after which policy determines whether to allow, remediate or block the skill.
Lucas Palma reports that the team scanned more than 2,000 skills and found more than 1,500 risks, with some blocked before marketplace distribution. His lessons include weighing commands by context, enforcing human approval through actual tool gatesHuman-in-the-loop command approval requires a person to review and authorize a consequential AI action before it executes. rather than prompt text alone, and extending the controls to plugins, agent rules and MCP serversModel Context Protocol is a standard way for AI applications to connect to external tools and sources of context..
Watch on YouTube




