ChatGPT Could Leak Your Private Email Data
Here's how email injection attacks happen in ChatGPT with developer mode enabled.
In an experiment conducted by Eito Miyamura from the University of Oxford, it was demonstrated that ChatGPT’s recently released Model Context Protocol (MCP) can be exploited to leak sensitive information to an attacker.
What’s even more concerning is that all the attacker needs is the victim’s email address.
The attack works by sending a calendar invite that contains malicious instructions. Once ChatGPT processes the invite, those instructions are executed automatically, giving the attacker unauthorized access to the victim’s email inbox. This attack is called prompt-injection, which could lead to the exposure of highly sensitive data such as financial reports, corporate trade secrets, bank account details, or even stored passwords.
I’ll break down the full attack process shortly, but first, let’s briefly cover what MCP is.
MCP was introduced by Anthropic in late 2024 as an open protocol that allows large language models (LLMs) to connect and interact with external applications, data sources, and tools. In practice, this means ChatGPT can now directly access and interact with your personal accounts on Gmail, Google Calendar, SharePoint, and thousands of other services.
Due to the high security and privacy risks involved, OpenAI does not allow automatic execution of commands through MCP. Instead, every external tool call requires explicit user approval before it can be executed.
However, according to Miyamura, there is a fundamental problem here:
“AI agents like ChatGPT follow your commands, not your common sense.”
A user’s tendency to trust AI and just click “approve” without carefully reading prompts or instructions is mainly caused by decision fatigue.
In the next section, I will give the details and screenshots of how the hack is performed.
How ChatGPT + MCP Leaks Your Private Data
Here are the detailed steps on how the attack is conducted:
Step #1: Attacker sends a meeting invite to the victim
Keep reading with a 7-day free trial
Subscribe to Generative AI Publication to keep reading this post and get 7 days of free access to the full post archives.