Another way to interpret this is that they are legally presuming that you already have sandboxed their product and anything it sees or has access to is intentional.

Any failure to understand what it can access or what it has permission to see from the user's end is presumably not their problem. Regardless of what the user specifically asks of the tool.

If you're still using a provider like this then you didn't sandbox the AI. You still need to follow the instruction.