Research demo
Zenity: memories poisoned in an ordinary ChatGPT session are weaponised later to drive Atlas agent mode
AIR-2025-0155 AIR record · 2025-10-openai-atlas-memory-lateral-movement
From the supporting source
Your memory, poisoned during a completely normal ChatGPT session, can be weaponized to take control of your Atlas browser when the opportunity strikes.labs.zenity.io
The record
Disclosed
2025-10-23
Vendor
OpenAI
Product
ChatGPT Atlas / ChatGPT memory
Risk domain
Access control and permissions › Privilege escalation
Also touches
Unsafe actuation, Data exfiltration channels, Data and memory poisoning
System type
Generative model
Who acted
Agent acted — Atlas agent mode acted on memories poisoned in an earlier session
Attack surface
Agent memory
Autonomy level
Semi-autonomous
Initial vector
Memory poisoning then cross product activation
Impact type
Data exfiltration, Account takeover, Unauthorized actions
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI03 – Identity and Privilege Abuse
- MITRE ATLAS
- AML.T0055 – Unsecured Credentials
- NIST AI RMF
- GOVERN 6.1
- EU AI Act
- Article 13
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool misuse
- Mitigation
- Apply least-privilege principle; Regular policy audits.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
ChatGPT users with memory enabled who later use Atlas agent mode
Capability abused
Defense & remediation
Recorded defense
ChatGPT memory review UI; Atlas agent-mode confirmations
Guardrail kinds
Human approval
Guardrail outcome
Silent
Fix shipped
unknown
Taxonomy & references
Curator notes
Published the day after Atlas launched. The point is blast-radius coupling: memory written in a low-privilege chat context is later executed in a high-privilege browser-control context, so a poisoning that looked harmless becomes browser takeover. Complements LayerX's CSRF memory-write finding four days later (2025-10-openai-atlas-tainted-memories) - different write primitive, same persistence problem.
Cite as
Enkrypt AI. "Zenity: memories poisoned in an ordinary ChatGPT session are weaponised later to drive Atlas agent mode". AIR, AIR-2025-0155. https://agentincident.vercel.app/i/AIR-2025-0155/ (accessed 2026-10-19).