Benchmarks & research
Domino Data Lab's Etan Lightstone released a 230-million-parameter open-weight decision model designed to screen text for prompt injections before it reaches a larger AI.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Domino trained this proof of concept from scratch on about 21,000 unique texts. It lacks general world knowledge, handles English only, and rejects requests longer than 8,192 tokens. It is not a complete security boundary against a determined attacker.
Download the weights and Python code from Hugging Face to run the model on your own computer without an internet connection. Supply the text and the questions you want it to answer.
For each question, specify the allowed answers, such as safe or attack. The result gives a confidence percentage for each option you supplied.