AK

Research Engineer — AI Evaluations & Agentic Safety

Publications

Public research outputs on empirical evaluation, agent memory, and agentic AI safety.

Utility Under Attack: Agent Memory Poisoning and the Limits of Content Screening and Provenance Ranking

This paper studies how a small amount of false information stored in persistent agent memory can sharply reduce downstream task accuracy. It evaluates content screening and provenance-weighted retrieval, identifies limits in both defenses under the measured LongMemEval setup, and motivates bounded occupancy constraints for retrieval. The paper reports the release of the evaluation harnesses, corpora, and aggregate run reports referenced in the study.

Plain-text citation

Karunanidhi, Arulnidhi. “Utility Under Attack: Agent Memory Poisoning and the Limits of Content Screening and Provenance Ranking.” arXiv preprint, arXiv:2608.21230, 2026. https://arxiv.org/abs/2608.21230