Evaluating using Mock Tool Calls to Quarantine Untrusted Prompt Inputs

David Gros, Adam Gleave

Large language models must frequently process untrusted inputs, such as judging an answer from another model or running tasks like spam and harm classifiers while under adversarial…