In AI platforms, skills are packaged capabilities that let an assistant retrieve knowledge or perform a business function, such as searching documents, filling forms, summarizing records, or triggering a workflow. A skill is more than a prompt template: it is a reusable platform layer with defined inputs, outputs, and often access to internal data or tools.
Skills matter because they move an assistant from conversation to action. If a model can select or invoke a skill based on untrusted text, prompt injection may redirect the system into using the wrong capability, exposing sensitive data or starting an unintended operation. Defenders should treat skills like privileged interfaces: apply least privilege, validate inputs, restrict which skills each assistant can reach, and isolate any skill that can touch internal services, files, or commands.