LLama 3.1 logs RAG gardrails issue

[trollkarlen]

When using your own logs in a RAG scenario the gardrails stoppes the llm to answer any questions due to Privacy. So maybe the gardrails is not needed when the information comes from the prompt itself and not from the training data.
Maybe the implementation makes this hard to differentiate ?

This works in llama3 but not in llama3.1