Show HN: Lelu – gate OpenAI agent actions on confidence and prompt injection (github.com)

Lelu has introduced an authorization engine specifically designed for AI agents, enhancing how actions and decisions are monitored and controlled. Unlike traditional authorization tools that merely prevent unauthorized access, Lelu addresses the critical issue of legitimate agents being misled through prompt injection or low-confidence decisions. Its significant capability lies in providing a human-in-the-loop monitoring system that flags questionable activities for review, ensuring safer AI operation in real-world applications. The engine employs a multi-layered pipeline that includes checks for shadow agents, prompt injection detection through various filters, and a confidence gating mechanism that evaluates decision reliability. Notably, it allows seamless integration with platforms like OpenAI and Anthropic, with a straightforward setup that can be executed in under a minute. This combination of robust security measures and user-friendly deployment is expected to resonate well within the AI/ML community, providing a much-needed safeguard against emerging threats in AI decision-making.