AI Potluck
Product / UX / Safety & Guardrails

Llama Prompt Guard 2

Meta

Meta's compact prompt-injection and jailbreak classifier (86M and 22M), labeling prompts benign or malicious. Part of LlamaFirewall, it is tuned for very low latency (sub-100ms) and high precision, with multilingual coverage across eight languages.

Llama Prompt Guard 2 comes in two sizes, 86M and 22M, the smaller trading a little recall for roughly a quarter of the latency. The card publishes no training data. The weights are behind Meta's access request rather than open to download. Verified 2026-08-13 via the HF model card.

Openness

3 medium confidence
3.0
weights
open(Llama-Prompt-Guard-2-86M/22M on HF)
data
not-released
license
Llama-4-Community(NOT OSI)

Open weights under Meta's Llama 4 Community License, which is use-restricted and not OSI-approved, so open_weights (3) like the rest of the Llama family.

Adoption

3 high confidence
3.0

111,334 downloads in the trailing 30 days for the one declared artifact, meta-llama/Llama-Prompt-Guard-2-86M, which bands at 100K-1M, level 3 on the software and model adoption scale.

Capability

4 medium confidence
4.0

Specialized injection/jailbreak classifier with strong accuracy and very low latency; complements full-taxonomy guards.

  • https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M recorded 2026-08-13

    Card evaluation table gives Llama Prompt Guard 2 86M an English AUC of .998, recall at 1% FPR of 97.5%, multilingual AUC .995 and 92.4 ms per classification on an A100; the 22M variant is listed at .995/88.7% and 19.3 ms. Repo is tagged 8 languages, and the card frames the task as prompt injection and jailbreak detection.

Verified 2026-08-13