New MemGhost Attack Plants Persistent False Memories in AI Agents Through One Email

This report summarizes a research paper demonstrating "stealth memory injection," an email-based technique (implemented by a tool called MemGhost) that can cause personal AI assistants which read email and write persistent memory to silently store false information; lab tests showed high success rates in background-mode runs against several agent frameworks. The authors document how the attacks evade input filters and hardened models, outline related prior incidents (EchoLeak, SpAIware), and recommend mitigations such as provenance tagging, user confirmation and separating untrusted email readers from memory-writing agents.