MemGhost Attack Lets Hidden Prompts Permanently Alter AI Agent Memory

A newly identified attack technique called MemGhost can manipulate AI agents into rewriting their own long-term memory using nothing more than a hidden instruction embedded in content like an email. Unlike traditional prompt injection, which tricks an agent into a one-time action, MemGhost causes the agent to permanently "remember" false information, such as an inflated transfer limit. Researchers report an 87.5% success rate against leading AI models, raising serious concerns about the security of persistent memory features that agent developers are rapidly deploying. The attack bypasses input filters by exploiting the agent's own legitimate memory-write tools rather than breaking any authorization boundary. Security experts argue the real fix lies in applying strict authorization controls and provenance checks to memory-write actions, rather than relying on prompt filtering alone.