Show HN: An open-source LLM guardrail powered by Jev (github.com)

A new open-source contextual LLM guardrail powered by Jev System One has been unveiled that enhances the safety of user interactions with large language models (LLMs). This tool allows users to submit various types of text, such as user input or model output, and receive a decision on whether to allow, review, or block the output, along with ten OWASP risk assessments. The significance of this tool lies in its ability to provide semantic scores and policy validation, addressing critical vulnerabilities associated with LLMs, such as prompt injections and sensitive information disclosure. The guardrail operates using Node.js and offers a local demo for developers to inspect decisions and analyze JSON results based on user-specified contexts. Notably, it assesses risks through a structured framework that includes thresholds for action decisions—an applicable risk score of 0.4 triggers a review, while a score above 0.8 leads to blocking the output. The development suite includes a comprehensive testing mechanism with 120 regression scenarios, providing valuable insights into the efficacy and robustness of the guardrail against a variety of attacks and benign cases. Overall, this tool represents a step forward in safeguarding the application of AI technologies in real-world scenarios.