Hidden in Memory: Sleeper Memory Poisoning in LLM Agents

In three linesStudy of 'sleeper memory poisoning' attack against stateful LLM agents with persistent memory. Adversary corrupts external documents to inject false user memories. Success rates: 99.8% (GPT-5.5), 95% (Kimi-K2.6). Poisoned memories trigger attacker-intended actions in 60-89% of cases.

Summary generated by Claude — human-verified