Ask HN: What is the most overlooked risk in the AI security domain?

A user is seeking information on overlooked risks in the AI security domain, particularly vulnerabilities related to AI, machine learning, and large language models that may be difficult to detect using typical pentesting methods. The user is interested in understanding what kinds of flaws are harder to catch and how pentesters will need to adapt to identify them. The question is focused on vulnerabilities beyond common issues like prompt injection.