Developer creates open-source runtime security system to protect LLM-powered agents from injection attacks, tool abuse, and data exfiltration in production environments.

System acts as a control plane around LLM pipelines, detecting and blocking prompt injections including obfuscation attempts in real time

Enforces structured tool calls with policy validation and prevents unauthorized data leakage of PII and sensitive information

Architecture includes gateway layer, context inspection, Open Policy Agent for policies, runtime enforcement with sandboxing, and Apache Kafka/Flink streaming pipeline

Addresses critical gap in production LLM systems where models have tool access and decision-making capabilities but lack dedicated security controls