OWASP Agent Memory Guard: Open-source runtime defense against AI agent memory poisoning attacks

Summary

OWASP's Agent Memory Guard is an open-source runtime defense layer that protects AI agents from memory-based attacks. It sits between an agent and its memory store, screening every read and write through a pipeline of detectors. The tool addresses a critical vulnerability where attackers can plant malicious text in an agent's persistent memory (conversation history, vector stores, scratchpads, RAG indexes) to override instructions, extract user data, or manipulate future tool calls. The defense achieves 92.5% recall at just 59µs latency, making it practical for real-time use.