Memory Poisoning: The AI Agent Attack Vector Nobody Is Scanning For

Short summary

Memory poisoning is a critical AI agent vulnerability where untrusted data written to vector databases or agent memory stores persists across all subsequent queries, unlike prompt injection. This attack vector isn't covered by OWASP ASI Top 10 and requires detection of 26+ memory sink patterns across vector databases, LangChain, and agent frameworks. The author released dfx-agentguard, detecting these patterns with 100% accuracy on 36 benchmarks.

• •Memory poisoning: untrusted data persists in vector databases, affecting all future agent decisions
• •Not covered by OWASP ASI Top 10; requires different detection than prompt injection
• •dfx-agentguard detects 26+ memory sink patterns across ChromaDB, Pinecone, Qdrant, LangChain, CrewAI, and AutoGen

Generated with AI, which can make mistakes.