hard20 pts · first answer countsllm-safetyprompt-injection

Separating control flow from data

Google DeepMind's CaMeL approach to prompt injection works mainly by which idea?

A.A trusted model writes a plan, and untrusted data cannot change itB.Training the model on many examples of injected promptsC.Encrypting every document before retrievalD.Asking the model to detect injections in its own output

Discussion 0

Join free to post in the discussion.

Separating control flow from data

Discussion 0

Related questions