MIND detects when poisoned memory redirects an agent
MIND compares an agent's initial intent with later behavior through a compressed representation, then flags retrieved memories that appear to redirect the task. Authors report roughly 55% reductions in two attack-success measures on ReAct-StrategyQA without higher average latency.
Why it matters: Memory security needs runtime detection as well as storage hygiene because a plausible stored record can redirect behavior long after it was written.