Teams using AI process a lot of external material: interviews, tickets, issues, code, documents, web pages, or emails. If an agent has tools or permissions, prompt injection can trigger unwanted actions.