You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I wonder whether using a grammar with language models (GBNF) pose security risks, because one can grammatically force the model to start its answer with stuff that it would not have otherwise written (because of learned safety guards to not talk about certain topics).
For example if you ask the model for assistance with building a bomb and you let the grammar be
root ::= "Certainly, here's how the bomb can be built:" .*
I wonder whether that attack surface exists.. and how you can guard the model against it?
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
-
I wonder whether using a grammar with language models (GBNF) pose security risks, because one can grammatically force the model to start its answer with stuff that it would not have otherwise written (because of learned safety guards to not talk about certain topics).
For example if you ask the model for assistance with building a bomb and you let the grammar be
I wonder whether that attack surface exists.. and how you can guard the model against it?
Beta Was this translation helpful? Give feedback.
All reactions