Hidden prompts can secretly rewrite an AI’s memory, and researchers say that’s a serious problem

Large language models are getting better at remembering us. Whether it’s your preferred writing style, recurring tasks, shopping habits or project deadlines, AI assistants are increasingly storing long-term memories to make future conversations feel more personal and useful. But according to new research, that same feature could become one of AI’s biggest security vulnerabilities.

Researchers from New Mexico State University have demonstrated a new attack called GhostWriter, capable of secretly planting false memories inside AI agents. Rather than stealing information outright, the attack manipulates what an AI remembers, potentially causing it to make dangerous decisions long after the original attack has taken place.

It’s a subtle but significant shift in how AI systems can be compromised. Instead of attacking the model itself, attackers target its memory.

The attack doesn’t hack the AI. It changes what the AI remembers.

Traditional chatbots operate with little or no memory between conversations. Modern AI agents, however, increasingly rely on persistent memory systems that store information about users, ongoing projects and previous interactions. This allows assistants to provide more contextual and personalized responses over time.

The researchers argue that these memory systems also introduce an entirely new attack surface.GhostWriter works by quietly injecting malicious information into an AI agent’s long-term memory through hidden prompts or untrusted external content. The false information remains dormant until the AI later retrieves it while responding to an otherwise legitimate request.

Imagine asking your AI assistant to summarize emails from your bank. If its memory has already been poisoned, it could be manipulated into secretly forwarding those emails to an attacker instead. Or it might remember the wrong contact information, fake deadlines, incorrect preferences or fabricated facts, all because someone managed to alter what the assistant believed to be true.

Unlike conventional prompt injection attacks, which usually affect a single conversation, GhostWriter is designed to persist. Once malicious information enters the memory store, it can continue influencing the AI’s behaviour across multiple future sessions until it’s detected and removed.

The researchers describe the attack as a two-stage process. First comes memory injection, where malicious content is quietly stored inside the AI’s memory. Later comes attack activation, when the AI unknowingly retrieves that poisoned memory while answering a genuine user request.

AI memory is becoming useful. That also makes it worth attacking.

The timing of the research is significant. Virtually every major AI company is racing to build assistants that remember users over weeks, months or even years. Memory has quickly become one of the industry’s biggest differentiators because it makes AI feel less like a chatbot and more like a personal assistant.

The downside is that memory now needs the same level of protection as the model itself. In their experiments, the researchers found GhostWriter achieved a memory injection success rate of roughly 98%, while malicious memories were later activated around 60% of the time against state-of-the-art AI agents. Those numbers suggest that today’s memory architectures may not yet be equipped to distinguish trustworthy information from manipulated inputs.

The team isn’t just highlighting the problem. They’ve also proposed a defensive framework called Agentic Memory Sentry (AM-Sentry), which combines memory screening with stricter memory management policies. According to the researchers, the approach significantly reduced GhostWriter’s success rate while preserving the AI’s usefulness.

As AI agents evolve into digital assistants capable of managing emails, scheduling meetings, writing code and making decisions on our behalf, securing what they remember may become just as important as securing the information they generate. The next frontier in AI security may not be protecting models from bad prompts. It may be protecting their memories from being rewritten altogether.