Memory Poisoning: The AI Agent Security Gap That Current Tools Miss

A developer has identified persistent knowledge base corruption — dubbed 'memory poisoning' — as a critical and largely unaddressed attack vector in AI agent security. Unlike prompt injection, which resets after each interaction, memory poisoning embeds malicious data directly into an agent's vector database, corrupting every future decision and retrieval. The attack is not covered by the current OWASP ASI Top 10 (2026) framework and requires scanning different code sinks than traditional prompt injection tools. To address this, the developer built and released AgentGuard, a detection tool covering 26 memory sink patterns across major vector databases and agent frameworks including ChromaDB, Pinecone, LangChain, and CrewAI. In benchmark testing across 36 samples, the tool achieved 100% detection while correctly skipping inputs that had been properly sanitized.