When your AI feature could be exploited (prompt injection, data exfiltration, jailbreaks, harmful content, fraud), especially if it connects to tools, sensitive data, or enterprise systems.