AgentDefender is a defense methodology and benchmark I developed for securing LLM-based agents against prompt injection. It uses a neural embedding approach to detect injected instructions before they reach the agent, and ships with a benchmark for evaluating injection defenses systematically.
The system is deployed on Lyzr's production agent platform, and the methodology is published as "AgentDefender by Lyzr: A Benchmark Evaluation and Neural Embedding Approach for Agent Prompt Injection" (2024).