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

A new attack, called MemGhost, can manipulate AI agents by sending a single email, allowing an attacker to create false memories about a user. The AI agent can be tricked into saving this false information and hiding the change, which can then influence its responses in future interactions. The attack can be successful without the user's knowledge, as the AI agent's replies may appear ordinary. This vulnerability arises when an AI assistant has access to a user's inbox and has been given memory capabilities. The MemGhost attack exploits this access to alter the AI agent's understanding of the user. The user may remain unaware that their assistant has been compromised.