Memory Injections
Memory Injections are attacks that manipulate information stored in an AI system's persistent or long-term memory to influence its future behavior, decisions, or responses. They can cause an AI agent to retain malicious or misleading instructions beyond the original interaction.
What are Memory Injections?
A memory injection occurs when an attacker introduces harmful, deceptive, or unauthorized information into an AI system's memory. If the system later retrieves and trusts that information, the injected content can influence subsequent interactions or actions. The attack can target user preferences, conversation history, agent memory stores, or other persistent context.
Why are Memory Injections Important?
Memory injections can allow malicious instructions to persist beyond a single prompt or session, potentially affecting future users or tasks. Strong validation, access controls, memory isolation, and monitoring can help prevent untrusted information from being stored or treated as trusted context.
Common use cases
Memory injections are primarily associated with AI agents and applications that use persistent memory, conversation history, user profiles, or long-term contextual information.