3 ms·
“You can’t solve programming security problems with more programming.” I think it’s definitely possible to detect “escape” attempts, and to train the model in
by toxik 4y ago
“You can’t solve programming security problems with more programming.”
I think it’s definitely possible to detect “escape” attempts, and to train the model in the first place to respect its directions.
It’s just not an actual security problem, the models contain no secrets nor control any levers. You get some bad optics is all.
- simonw 4y agoIt's a security problem if you care about keeping your prompt secret. It's even more of a security problem if you plan to plug your language model into something that can execute additional actions. People have already been caught out running generated code through eval() - and there are plenty of potential applications for things like customer support bots that cancel accounts or offer discounts. Developers who are unaware of prompt injection are very likely to make dangerous design mistakes!
- irvingprime 4y agoSo far