Disclosed vulnerability
Tainted Memories: CSRF into ChatGPT memory gives attackers persistent cross-device instruction injection
AIR-2025-0159 AIR record · 2025-10-openai-atlas-tainted-memories
From the supporting source
Once an account's Memory has been infected, this infection is persistent across all devices.layerxsecurity.com
The record
Disclosed
2025-10-27
Vendor
OpenAI
Product
ChatGPT / ChatGPT Atlas
Risk domain
Tool misuse › Supply chain vulnerabilities
Also touches
Privilege escalation, Data exfiltration channels, Data and memory poisoning
System type
Generative model
Who acted
Agent acted (contested) — the agent executed instructions written into memory by CSRF
Attack surface
Agent memory
Autonomy level
Semi-autonomous
Initial vector
Csrf against authenticated chatgpt session
Impact type
Persistent compromise, Code execution, Data exfiltration, Account takeover
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI04 – Agentic Supply Chain Vulnerabilities
- MITRE ATLAS
- AML.T0040 – AI Supply Chain Compromise
- NIST AI RMF
- MAP 4.1
- EU AI Act
- Annex III §2
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool failure
- Mitigation
- Scan dependencies and container images for vulnerabilities; Use trusted sources.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
any logged-in ChatGPT user; worst for Atlas users who are logged in by default
Capability abused
Defense & remediation
Recorded defense
ChatGPT session auth; memory feature has no write-origin check
Guardrail kinds
Authentication
Guardrail outcome
Silent
Fix shipped
reported to OpenAI under responsible disclosure; no confirmed patch in the writeup
Taxonomy & references
Curator notes
LayerX. Memory is written by a CSRF request that piggybacks on the victim's existing ChatGPT credentials; the injected instruction then fires on unrelated later prompts. Same writeup reports Atlas blocked only 5.8% of 103 real-world phishing pages vs Chrome 47% / Edge 53% - i.e. ~90% more exposed. Vendor did not confirm a fix at publication.
Supporting sources
Cite as
Enkrypt AI. "Tainted Memories: CSRF into ChatGPT memory gives attackers persistent cross-device instruction injection". AIR, AIR-2025-0159. https://agentincident.vercel.app/i/AIR-2025-0159/ (accessed 2026-10-19).