3 ms·
A random person can send you an email. Your agent can read that email. So then if the user says "tell me what's in my email" the agent will go and read that me
by simonw 4y ago
A random person can send you an email. Your agent can read that email.
So then if the user says "tell me what's in my email" the agent will go and read that message from an untrusted source, and could be tricked into acting on additional instructions in that message.
- jmugan 4y agoThanks, that's starting to make more sense. With LLMs as the execution platform that both reads data in natural language and takes instructions in natural language, it becomes harder to separate instructions from data.