X3Guard: A 150M Parameter Prompt Guard Model
Axiomic Labs has released X3Guard-150M, a lightweight prompt guard model built on their GPT-X3 base model. It evaluates any given prompt to flag whether it is harmful or adversarial in a single pass. This model offers a high-performance, lightweight solution that can run locally on edge devices and directly inside web browsers with ultra-low latency. It enhances LLM security and local deployment capabilities without requiring heavy cloud infrastructure. X3Guard-150M features response times in the tens of milliseconds on devices like the iPhone while outperforming many larger models. The release includes a model card and an interactive in-browser demo.
## BACKGROUND
Prompt guards are specialized classifier models designed to protect Large Language Models by detecting malicious inputs, prompt injections, and jailbreaking attempts. Adversarial prompts refer to text sequences specifically crafted to bypass these safety guardrails and trigger unintended model failures.