🛡 Lelu: Protecting AI Agents from Runtime Manipulation
Lelu has been released—an open-source authorization engine for AI agents designed to protect against runtime manipulation (e.g., prompt injection). The system utilizes a multi-layered pipeline that includes injection filtering, a confidence gate based on LLM token probabilities, and risk assessment.
🌍 It addresses the security challenges of autonomous agents, where standard IAM tools fail to recognize breaches caused by context manipulation.
👤 It allows developers to build agents with a "safety fuse" that analyzes the appropriateness of AI actions.
Source 1: https://github.com/lelu-auth/lelu