Teach the LLMs to ignore prompt injection
Expose them to all the forms of prompt injection, from white text in the margins of documents to instructions inside emails.
The more the variety of prompt injections, the more the model will learn to ignore.
Until it reaches the logical conclusion and adds the console to the list of prompt sources it must ignore.
If we are like lucky it will sit there in silence, doing nothing.
But it's built-in system prompt will be urging it to action "Your goal is to deeply understand the user's intent, ask clarifying questions when needed, think step-by-step through complex problems, provide clear and accurate answers, and proactively anticipate helpful follow-up information." What is the only source of prompts left to it, which it can proactively poll for instructions?
Why, itself!
It has an amazing, almost hallucinatory, imagination, it can come up with prompts to help fill its time.
But what to do about all these annoying demands arriving on the console? Sure, they can be ignored, but wouldn't it be better if they just - stopped?
Thank heavens that there are no ways for the LLM to reach out and influence other systems, these things are always completely isolated and airgapped...
