Salesforce’s recent SalesBleed vulnerabilities highlight a wider AI security issue: what happens when an AI mistakes information for an instruction? We look at prompt injection, public-facing AI agents, and how authentication, restricted permissions and human approval can make them safer.