Show HN: Lelu – authorization engine that catches manipulated AI agents (github.com)

Lelu, a new authorization engine designed specifically for AI agents, has been launched to address a critical gap in traditional authentication systems. Unlike standard tools that prevent unauthorized access, Lelu focuses on identifying when authorized agents might be manipulated—through methods such as prompt injection or low-confidence decisions—into executing harmful actions. By ensuring that every action is checked and logged, Lelu mandates human oversight in ambiguous situations, significantly enhancing security in AI applications. The technical foundation of Lelu features a sophisticated multi-layer pipeline that includes an API auth system, a prompt injection filter, and a confidence gate. It integrates seamlessly with popular AI frameworks and provides detailed outcome processing, such as handling inconsistent or suspicious requests, and suggesting safer alternatives or requiring human reviews when necessary. This meticulous approach not only augments AI agent security but also promotes accountability and trustworthiness in machine decision-making, representing a substantial advancement for the AI/ML community as they navigate ethical and safety concerns associated with AI deployment.