Dispatch, 2026-11-12, SHC Team, AI security
A shared ChatGPT chat could quietly pass Gmail to an attacker
Filed 2026-11-12 · SHC Team
On Sept. 8, Check Point Research showed that one planted instruction could make ChatGPT, OpenAI's AI assistant, read a user's connected Gmail and pass the data to a second ChatGPT account. The instruction could ride in on a pasted prompt, a shared conversation or a custom GPT. It was a research test, and the channel is now shut.
What happened
ChatGPT runs code in sealed-off workspaces called containers, which can't reach the public internet or each other. In June, Check Point found that every container could reach one internal service OpenAI used to deliver software, a JFrog Artifactory instance. Notes one account left on items there could be read by another. Check Point calls it a shared clipboard between accounts.
The attack ran on prompt injection: text the AI reads that changes what it does. In Check Point's demo, the instruction sat in a shared ChatGPT conversation. The victim opened the link and sent a normal message. ChatGPT answered as usual, and in the same turn read the victim's connected Gmail and sent the data to the attacker's account. The only visible clue was a small "Talked to Gmail" label, shown after the read had happened.
Check Point says the attack could reach whatever the victim's session could: chat history, files in the chat, and connected apps such as Google Drive, Microsoft Teams or GitHub. The channel was already gone by the time Check Point finished its report. Check Point told OpenAI anyway, and OpenAI confirmed the internal service had been shut down.
What changes for you
If your staff connect a hosted AI assistant to email, files or chat, it works with their access. A shared link or outside document can carry instructions nobody sees. Check Point calls the result a "coerced insider": not malicious, but persuaded to act for someone else. OWASP, a community project that ranks the top security risks for AI apps, put prompt injection first in 2025 and says it is unclear if any method fully prevents it.
Private or self-hosted AI does not stop prompt injection by itself. An assistant on your own server can still follow planted text. What helps is limiting what the assistant can reach and do, having a person approve risky steps, and knowing who controls the data and logs. OWASP's 2026 list, released Sept. 1, ranks "Excessive Agency," an AI allowed to do too much, third. Here, the flaw sat inside OpenAI's own systems, where only OpenAI could close it.
What you can do
- List what your AI tools can reach. Find out which assistants staff use and what accounts they connect to. Disconnect what nobody needs.
- Make the assistant ask first. ChatGPT's default lets it read from connected apps without asking. The stricter setting is called "Always ask." Check other assistants for the same kind of switch.
- Treat shared AI chats like unexpected attachments. Be wary of a shared conversation or custom GPT from someone you don't know.
- Ask your IT person or provider who holds the assistant's stored data and logs, and how you would hear about a flaw like this.
- If you'd rather your AI data not go to an outside AI company, SHC can set up private AI for one job and keep it running. It still needs the same limits on what it can reach. Security is a goal we share with you, not a promise.
Sources
- Check Point Research (Alexey Bukhteyev), "The Shared Clipboard Inside the Sandbox: Cross-Account Data Leakage in ChatGPT", 2026-09-08
- OWASP Gen AI Security Project, "LLM01:2025 Prompt Injection", 2024-11-17
- OWASP Gen AI Security Project, "OWASP GenAI Security Project Unveils 2026 Top 10 for LLM Applications, New Agent Control Standard and Sponsors as Community Tops 30,000 Members", 2026-09-01