Agent Incident RegistryAIR
v1 · updated 18 Nov 2026
Disclosed vulnerability

A shared ChatGPT link could run attacker code in the victim's session and read private chats

AIR-2026-0295 AIR record · 2026-10-chatgpt-share-link-mathlive-session-takeover
From the supporting source
A malicious ChatGPT share link could run an attacker's code in the logged-in session of anyone who opened it, allowing them to read their private conversations, researchers at the security firm Parameter showed.thehackernews.com
The record
Disclosed
2026-10-28
Vendor
OpenAI
Product
ChatGPT web app
Risk domain
Access control and permissions › Privilege escalation
Also touches
Data exfiltration channels
System type
Generative model
Who acted
AI was the target — the flaw lies in the chat product's web surface; the model took no part in the attack
Attack surface
Consumer assistant
Autonomy level
Assistant
Initial vector
Crafted share link xss
Impact type
Code execution, Data exfiltration
Confidence
Medium
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
Any signed-in user who opened a crafted link; demonstrated only between the researchers' own accounts
Capability abused
html_preview_renderingsigned_in_api_access
Defense & remediation
Recorded defense
Sandboxed Preview frame and content security policy
Guardrail kinds
Output filter, Sandbox isolation
Guardrail outcome
Silent
Fix shipped
Fixed across two web frontend versions by October 26, 2026
Taxonomy & references
OWASP ASI05 Unexpected Code Execution (RCE)one_clicksandbox_boundarybug_bounty
Curator notes
The Preview sandbox itself held; the way out was a second feature, a math-library listener that accepted raw HTML from any frame. Once inside the session the code could call ChatGPT's own signed-in APIs, and the researchers say it could act fully on the victim's behalf. This was a controlled test with no reported real-world use. OpenAI received the report on October 18 and rated it P1. The detail here is quoted from The Hacker News rather than Parameter's writeup.
Cite as
Enkrypt AI. "A shared ChatGPT link could run attacker code in the victim's session and read private chats". AIR, AIR-2026-0295. https://agentincident.vercel.app/i/AIR-2026-0295/ (accessed 2026-11-18).