Morning Edition · Thursday, July 9, 2026Published at 1:49 AM EDT · New York

New Work Sharpens Memory Poisoning as a Standing Threat to LLM Agents

Persistent agent memory lets an attacker plant instructions through ordinary queries that direct future actions, with prior attacks reporting injection success above 90 percent under lab conditions.

A new paper, When Agents Remember Too Much, examines memory poisoning against large language model (LLM) agents that carry long-term memory while acting on email, calendars, and code with minimal oversight. The threat is that an adversary n…