MIND: Lightweight Memory Injection Defense for LLM Agents via Intent-Aware Information Bottleneck

The paper introduces MIND, a lightweight defense against memory injection attacks in memory-augmented LLM agents. It uses an intent-aware Information Bottleneck to derive compact representations linking the initial user intent with turn-level behavior, filtering repetitive or task-irrelevant context before a lightweight malicious-memory detector is applied. On ReAct-StrategyQA, the authors report reductions of 55.4% in mean ASR-r and 55.3% in mean ASR-a, while matching the undefended agent in average task accuracy and latency. The reported evidence is currently centered on the abstract and named benchmark.