OWASP Agent Memory Guard Released — Open-Source Runtime Defense for Agentic Memory Poisoning (ASI06)

OWASP released Agent Memory Guard, the reference implementation for OWASP ASI06 (Memory Poisoning) in the OWASP Top 10 for Agentic Applications. The open-source tool sits between an agent and its memory store, screening every read/write through a pipeline of five detectors (prompt injection, secret/PII leakage, protected-key tampering, size anomaly, SHA-256 integrity baselines) and a YAML policy with allow/redact/quarantine/block actions. It ships as a drop-in wrapper for LangChain, LlamaIndex, and CrewAI, with 59-microsecond median latency. Benchmark results: 92.5% recall, 100% precision, 0 false positives across 55 test cases.

Agent memory poisoning is a qualitatively different threat from prompt injection: a malicious instruction written to long-term memory persists across sessions and can trigger weeks later as trusted context, bypassing both input sanitization and output validation. OWASP Agent Memory Guard is the first production-grade open-source defense layer for this attack class, making it immediately actionable for teams running LangChain or LlamaIndex agents with persistent memory stores.

Any team running LangChain/LlamaIndex/CrewAI agents with ConversationBufferMemory, Redis chat histories, vector stores, or RAG indexes should evaluate this as a low-friction defense layer. The 3-line integration pattern (pip install + wrap existing backend) means security teams can add coverage without refactoring agent code. Higher-threat-model deployments should layer additional detection on top of the open-source base.