prompt-guard

Load the meta-llama/Prompt-Guard-86M model with transformers and score each input's jailbreak probability from the softmax logits. Block inputs where the score exceeds your threshold, typically 0.5 for balanced security.

Prompt Guard detects jailbreaks and prompt injections in inputs, while LlamaGuard moderates content like violence or hate speech. They address different threats and are commonly combined for defense-in-depth.

Yes, tokenize documents in batches of 32 with padding and truncation, then sum the INJECTION and JAILBREAK probabilities per document. Use a lower threshold around 0.3 for third-party data filtering.

The model truncates inputs at 512 tokens, so injections near the end of long texts may be missed. Use a sliding window with overlapping chunks and take the maximum score across chunks.

Technical queries about prompt injections can trigger false positives because they resemble attack patterns. Raise the threshold for trusted users, for example from 0.5 to 0.7, to reduce false positives.

Prompt Guard supports eight languages: English, French, German, Spanish, Portuguese, Italian, Hindi, and Thai. It runs on CPU at 50-200ms latency or GPU at under 2ms per request.