Analysis
Memory Injection Attacks on LLM Agents via Query-Only Interaction
First reported · Discovered arxiv.org
Page published
Earliest dated coverage: 24 Oct 2026 · First observed: 27 Oct 2026 · Latest dated coverage: 24 Oct 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
Persistent agent memory can carry malicious influence into otherwise clean conversations, making memory creation and retrieval important security boundaries for defenders.
Memory poisoning can preserve attacker-influenced content across AI-agent sessions, as explained in Surajkhaitan’s overview of persistent prompt injection. The supporting MINJA study demonstrates how attackers can insert malicious memory records through queries and output observations without directly editing an agent’s memory bank.