Analysis · curated 8 Sep 2026
AI Agent Memory Poisoning: Attacks That Outlive the Session | Agentic AI Security 101 | Part 4 of 12
First reported youtube.com
Coverage timeline
Single-source analysis — first reported, latest, and curated coincide.
Why it matters
Memory poisoning defeats guardrails that only inspect current context because the malicious payload acts across sessions, users, and time, making it a persistence and exfiltration risk for any agent with long-term memory and document ingestion.
Part 4 of ZioSec's Agentic AI Security 101 series explains AI agent memory poisoning, where an injection planted in one session is written to long-term memory and triggered later in a different session or for a different user. The video cites Johann Rehberger's SpAIware (a persistent Copilot backdoor combining memory poisoning with automatic exfiltration) and Delayed Tool Invocation, and recommends controls like treating memory writes as privileged operations, tenant segmentation, memory quarantine, and rollback with lineage.