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

Researchers have identified a new attack technique called 'stealth memory injection,' dubbed MemGhost, that enables attackers to plant persistent false memories in AI agents through a single malicious email. The attack manipulates AI assistants that read emails and maintain persistent memory files, allowing an adversary to silently alter the agent's behavior in future sessions without detection. The technique was successfully tested against OpenClaw and Claude-based agents, with high success rates in background execution modes, highlighting a critical risk in AI systems that combine untrusted input processing with autonomous memory updates.

AI-extracted · verify before operational use