Llama Prompt Guard 2 86M

Llama Prompt Guard 2 86M is the larger variant of Meta's prompt injection detection model, offering higher accuracy for identifying malicious prompts and jailbreak attempts at the cost of slightly more compute. With 86 million parameters, it provides more nuanced classification of attack patterns.

The model detects a wide range of prompt injection techniques including direct injections, indirect injections through retrieved content, multi-turn manipulation attempts, and encoded/obfuscated attacks. It outputs probability scores for different threat categories.

Prompt Guard 2 86M is recommended when maximum detection accuracy is needed and the additional compute overhead is acceptable, such as in high-security applications.