prompt-guard

Detect prompt injections and jailbreak attempts by running text through a classifier that returns probabilistic scores for both unsafe classes. This allows you to gate malicious prompts before they reach your LLM.

Yes, you can deploy jailbreak detection on CPU using an 86M parameter compact model designed for low latency. This lightweight architecture suits chatbots and API endpoints without requiring dedicated GPU hardware.

Multilingual prompt injection detection supports 8 languages, enabling you to identify jailbreak attempts and injections across diverse user bases in global LLM applications.

Integrate jailbreak detection into a RAG pipeline by passing user queries through the classifier to obtain safe or unsafe verdicts. This gates unsafe prompts before they trigger retrieval or generation.

A prompt injection classifier returns probabilistic scores for INJECTION and JAILBREAK classes. These scores provide the threshold values needed to gate unsafe prompts or filter generated outputs.

No, you do not need heavy GPU resources because prompt safety detection runs on a lightweight 86M parameter model. It is optimized for on-device or CPU deployments with simple Python integration.