Comment by Aaron_NW

Great point. As we use agents to write and summarize responses, we need to treat the source content as a prompt injection surface.

Curious if you have a specific method for checking the content beyond the overall instruction not to trust it?