liveBook · Manning

Memory poisoning

As explained in AI Governance

Memory poisoning is the insertion of misleading or malicious information into an agent’s memory to influence its later behavior.

Memory poisoning inserts misleading or malicious information into an agent’s persistent memory. When another user or a future session retrieves it, the information can trigger an attack that outlasts the original conversation and compound over time. section 4.8.1

Persistent-memory attack

The ZombieAgent attack chain disclosed in January 2026 involved sending a victim a file containing hidden instructions intended to modify ChatGPT’s persistent memory. section 4.8.1

When investigating suspicious agent behavior, log reasoning traces—including planning steps, tool choices, and intermediate conclusions. These records can help determine whether memory poisoning influenced the agent. section 4.8.8

How AI Governance covers it

How deep it goes
Explains how malicious information can persist and activate when retrieved, contrasting this with conversation-bound prompt injection.
Examples
Uses the ZombieAgent attack chain against ChatGPT, involving hidden instructions in a file.
What you can do after
Diagnose possible memory poisoning by reviewing logged reasoning traces, planning steps, and tool selections.
How it connects
Links memory poisoning to persistent memory, prompt injection, and autonomous agents.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.