(on mobile so not going to bother with pasting references) The threat model is, IMHO, absolutely ridiculous and assumes that humans have already delegated all

(on mobile so not going to bother with pasting references) The threat model is, IMHO, absolutely ridiculous and assumes that humans have already delegated all tasks to an LLM. The mere requirement that _people read their own emails before sending them_ wipes out this threat entirely.