Show HN: ReasonGate- An explainable gate that blocks LLM prompt injection (github.com)

ReasonGate has been introduced as an explainable security gate designed to safeguard large language model (LLM) applications from prompt injection attacks, a prevalent vulnerability highlighted in the OWASP LLM Top 10. By placing this gate before the model, it ensures that potentially malicious instructions embedded within user prompts or retrieved document contexts can be identified and blocked before they reach the LLM, significantly mitigating the risks of data breaches and unauthorized actions. This step is crucial as it replaces traditional black-box security measures with transparency, providing detailed audits on which signals triggered blocks and the reasoning behind these decisions. The technical framework of ReasonGate is noteworthy, as it employs a multi-layered approach that includes normalization, injection detection, and leakage prevention, all while being model-agnostic, meaning it can interface with any LLM configuration. The core is built in pure Python with zero dependencies, allowing for easy installation and operation within isolated environments. ReasonGate’s unique design allows for a structured JSON output, offering comprehensive audit trails for compliance requirements. With robust performance metrics, including high recall rates and established methods for handling obfuscated evasion tactics, ReasonGate promises to enhance the security posture of LLM applications, ensuring that developers can deploy AI solutions without compromising safety and accountability.