Agents with long-term memory or retrieval can be poisoned once and stay compromised across sessions, because planted instructions or false facts can be pulled back into context later. Attacks target chat memories, RAG knowledge bases, and shared agent state.
Read in this order
- ArticleChatGPT: Hacking Memories with Prompt InjectionStart here: a clear demo of planting false memories through injected content.
- PaperAgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge BasesHow poisoning an agent's memory or knowledge base steers later behavior.
- ArticleManipulating AI memory for profit: The rise of AI recommendation poisoning (Microsoft Security)Memory poisoning found in the wild, planted by real companies, February 2026.