This is a masterful demonstration of the unsecurable nature of LLMs, and how prompt injection by dedicated humans who know how to write will always win.

https://role-confusion.github.io/