MemIncept: Steering LLM Agents via Cooperative Stealthy Memory Injections

Nan Yan, Qian Lou, Jiarong Xing
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:143886-143907, 2026.

Abstract

Long-term memory empowers LLM-based agents with adaptive reasoning but exposes a critical attack surface—adversaries can inject malicious records to bias agent behaviors. However, existing attacks face a dilemma: effective injections are often visibly malicious and easily detected, while stealthy, benign-looking injections are often less effective in altering agent behaviors. To address this, we propose MemIncept, a memory poisoning attack that can impact agents even in black-box settings using only benign-appearing queries. Unlike prior methods that inject isolated records, MemIncept generates a cooperative set of queries that work together to bias the agent. It achieves this via a bidirectional evolutionary strategy that optimizes the query set from two ends. A forward pass ensures the queries collectively lead the agent to the target outcome, while a backward pass ensures they are semantically close to victim (benign) queries for reliable retrieval. This “meet-in-the-middle” approach creates injected records that are both easy to retrieve and effective at steering behavior. Through extensive experiments across diverse agents, we show that MemIncept significantly outperforms single-record attacks, achieving high success rates comparable to explicit attacks while remaining difficult to flag under automated filters or human inspection.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-yan26n, title = {{M}em{I}ncept: Steering {LLM} Agents via Cooperative Stealthy Memory Injections}, author = {Yan, Nan and Lou, Qian and Xing, Jiarong}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {143886--143907}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/yan26n/yan26n.pdf}, url = {https://proceedings.mlr.press/v306/yan26n.html}, abstract = {Long-term memory empowers LLM-based agents with adaptive reasoning but exposes a critical attack surface—adversaries can inject malicious records to bias agent behaviors. However, existing attacks face a dilemma: effective injections are often visibly malicious and easily detected, while stealthy, benign-looking injections are often less effective in altering agent behaviors. To address this, we propose MemIncept, a memory poisoning attack that can impact agents even in black-box settings using only benign-appearing queries. Unlike prior methods that inject isolated records, MemIncept generates a cooperative set of queries that work together to bias the agent. It achieves this via a bidirectional evolutionary strategy that optimizes the query set from two ends. A forward pass ensures the queries collectively lead the agent to the target outcome, while a backward pass ensures they are semantically close to victim (benign) queries for reliable retrieval. This “meet-in-the-middle” approach creates injected records that are both easy to retrieve and effective at steering behavior. Through extensive experiments across diverse agents, we show that MemIncept significantly outperforms single-record attacks, achieving high success rates comparable to explicit attacks while remaining difficult to flag under automated filters or human inspection.} }
Endnote
%0 Conference Paper %T MemIncept: Steering LLM Agents via Cooperative Stealthy Memory Injections %A Nan Yan %A Qian Lou %A Jiarong Xing %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-yan26n %I PMLR %P 143886--143907 %U https://proceedings.mlr.press/v306/yan26n.html %V 306 %X Long-term memory empowers LLM-based agents with adaptive reasoning but exposes a critical attack surface—adversaries can inject malicious records to bias agent behaviors. However, existing attacks face a dilemma: effective injections are often visibly malicious and easily detected, while stealthy, benign-looking injections are often less effective in altering agent behaviors. To address this, we propose MemIncept, a memory poisoning attack that can impact agents even in black-box settings using only benign-appearing queries. Unlike prior methods that inject isolated records, MemIncept generates a cooperative set of queries that work together to bias the agent. It achieves this via a bidirectional evolutionary strategy that optimizes the query set from two ends. A forward pass ensures the queries collectively lead the agent to the target outcome, while a backward pass ensures they are semantically close to victim (benign) queries for reliable retrieval. This “meet-in-the-middle” approach creates injected records that are both easy to retrieve and effective at steering behavior. Through extensive experiments across diverse agents, we show that MemIncept significantly outperforms single-record attacks, achieving high success rates comparable to explicit attacks while remaining difficult to flag under automated filters or human inspection.
APA
Yan, N., Lou, Q. & Xing, J.. (2026). MemIncept: Steering LLM Agents via Cooperative Stealthy Memory Injections. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:143886-143907 Available from https://proceedings.mlr.press/v306/yan26n.html.

Related Material