Memory Poisoning
An attack method targeting the memory mechanisms of AI agents, in which an attacker crafts inputs that write false information into an AI system's long-term memory storage. This persistently influences the system's subsequent interactions with all users. Memory poisoning represents an evolved form of prompt injection attacks, distinguished by its durable effect on stored knowledge rather than transient manipulation of a single session.
Timeline (last 90 days)
Memory poisoning is harder to detect than one-off prompt injection because the contamination signal is written long before the malicious behavior is triggered.
Memory poisoning does not occur in real-time conversation, so it does not trigger any real-time security alert, and may only be triggered days or even weeks later.
The fix for memory poisoning is memory versioning: tagging each memory with a timestamp and a validity period label.
Memory poisoning is an evolved form of prompt injection attack, in which an attacker induces an AI to write false information into long-term memory storage, thereby influencing the system's subsequent interactions with all users.
All Facts (4)
Memory poisoning is harder to detect than one-off prompt injection because the contamination signal is written long before the malicious behavior is triggered.
50%UnverifiedMemory poisoning does not occur in real-time conversation, so it does not trigger any real-time security alert, and may only be triggered days or even weeks later.
50%UnverifiedThe fix for memory poisoning is memory versioning: tagging each memory with a timestamp and a validity period label.
50%UnverifiedMemory poisoning is an evolved form of prompt injection attack, in which an attacker induces an AI to write false information into long-term memory storage, thereby influencing the system's subsequent interactions with all users.
50%