Agent Harness Tutorial · End to end · Zoom level 3
How we take care of it: Prompt Guard 2
Prompt Guard 2 catches paraphrased and foreign-language attacks.
How we took care of it.
- Not built yet. Llama Prompt Guard 2 is a small open classifier for injection attempts.
- It would plug in as an hf_classifier evaluator: one line in evaluators.yaml, no new port.
- It would catch what the phrase list misses: paraphrases and other languages.
| It reads | the user message, and later each tool result |
| Kind | small model |
| Provided by | open source |
| If it fails | safe reply / hold writes |
| Where in the code | planned · finding F-132 |
| Runs at | 2 · Before the model, 6 · After the tool |
Real example—