PromptGuard 2

A high-precision, multilingual classifier designed to intercept prompt injections and jailbreak attempts before they reach your LLM.

PromptGuard 2 is Meta's latest evolution in LLM security, offering a specialized defense layer built on the DeBERTa architecture. This update introduces two distinct models: a high-performance 86M parameter version and a lightweight 22M parameter variant that slashes compute costs by 75% for latency-sensitive applications. By utilizing a custom energy-based loss function and adversarial-resistant tokenization, the system effectively neutralizes complex threats like DAN-style jailbreaks and whitespace manipulation across eight major languages. It functions as a standalone firewall, allowing developers to filter malicious inputs at the edge of their pipeline without compromising the primary model's core utility.

https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M