Prompt injection and agent abuse rarely look like attacks. Learn how to detect AI misuse in production by monitoring behavior, not just prompts.