Defend agent memory against forged-reasoning poisoning, not just forged facts (FARMA/SENTINEL)

arXiv·high signal

A July 6 arXiv paper introduces FARMA, an attack that poisons an agent's stored reasoning traces (not factual entries), evading keyword filters and self-reinforcing to defeat consensus defenses — up to 100% success. The defense, SENTINEL, adds a 'Reasoning Guard' that structurally analyzes suspect memory entries with five weighted signals, cutting success to ~0% with no false positives across 326 benign traces. Actionable: audit long-term agent memory for self-referential reinforcement patterns and validate reasoning quality, not just facts.