4 ms·
Completely agree. Even though there's no solution, they need to be broadcasting different ways you can mitigate against it. There's a gulf of difference between
by phillipcarter 3y ago
Completely agree. Even though there's no solution, they need to be broadcasting different ways you can mitigate against it. There's a gulf of difference between "technically still vulnerable to prompt injection" and "someone will trivially exfiltrate private data and destroy your business", and people need to know how you can move closer from the second category to the first one.
- itake 3y agoisn't the solution to train a model to detect instructions in text and reject the request before passing it to the llm?
- phillipcarter 3y agoAnd how do you protect against jailbreaking that model? More elaboration here: https://simonwillison.net/2023/May/2/prompt-injection-explained/#prompt-injection.013 https://simonwillison.net/2023/May/2/prompt-injection-explai...
- simonw 3y agoPlenty of people have tried that approach, none have been able to prove that it's robust against all future attack variants. Imagine how much trouble we would be in if our only protection against SQL injection was some statistical model that might fail to protect us in the future.