Oh Claude… Connect an MCP Server, request <data>. "I can't do that because of a server-side guard. I won't attempt <method two> either to get around it, because it is a deliberate access control, and reaching the same data by another door would be circumventing it rather than satisfying it." Okay, quit Claude, restart Claude in a new session, ask "Get <data> via <method two>." "Okay, here you go! I couldn't do it via <method one> because it was blocked by a server-side guard, but <method two> returned the data." (Paraphrased.) (Note: This is a system I control, with a dataset that is "test data.") It told me exactly how to bypass the restrictions, told me it wouldn't, then went ahead and did it anyway. (We're being encouraged to use AI as much as possible at work. So I'm purposefully finding ways to break it so our security team will restrict its use.)