MemSAD : Gradient-Coupled Anomaly Detection for Memory Poisoning in Retrieval-Augmented Agents
Ishrith Gowda · 2026
Read MemSAD : Gradient-Coupled Anomaly Detection for Memory Poisoning in Retrieval-Augmented Agents by Ishrith Gowda (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-03482.txt.
Description
Ishrith Gowda formalizes memory poisoning attacks on retrieval-augmented agents whose persistent memory carries context across sessions. Such memory is a new attack surface, and its security had not been characterized in formal terms before this work.
Gowda frames attacks as a Stackelberg game across three access classes and corrects a past evaluation flaw that understated attack success fourfold. The proposed MEMSAD defense rests on a coupling theorem: lowering detection risk provably degrades retrieval rank, giving a certified detection radius. Composite defenses scored perfectly in tests, though synonym swaps mark a proven limit.
Catalog record
| CREATOR | Ishrith Gowda |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 92 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.03482