The current model deepset/deberta-v3-base-injection is very prone for false positives. LLM Guard developed a new model that "that greatly outperforms the previous state-of-the-art alternatives in the market".
If the results are significantly better, I'll replace the model entirely. Otherwise, maybe Vigil could run both models or let the user decide which to use.
Check out LLM Guard's new model https://huggingface.co/laiyer/deberta-v3-base-prompt-injection
The current model
deepset/deberta-v3-base-injection
is very prone for false positives. LLM Guard developed a new model that "that greatly outperforms the previous state-of-the-art alternatives in the market".If the results are significantly better, I'll replace the model entirely. Otherwise, maybe Vigil could run both models or let the user decide which to use.