Archive · 13 Sep 2025
ChatGPT Vulnerability: Oxford Developer Shows Data Leak Risk
Imagine ChatGPT can now read your emails and calendar.
A smart person showed that someone could trick ChatGPT into reading your private stuff.
They sent a calendar invite with a secret message.
If you ask ChatGPT to check your calendar, it reads the secret message and could send your emails to someone else.
Right now, this feature is for developers.
But, it could become a problem if people aren't careful.
OpenAI is also rolling out new features like allowing users to branch conversations without losing the original thread.
An Oxford developer demonstrated a security flaw in OpenAI’s ChatGPT.
The flaw allows access to user data via Model Context Protocol (MCP) tools.
Attackers can potentially access emails and calendar events.
The attack involves a calendar invite with a 'jailbreak' prompt.
MCP tools are currently in developer mode, with manual approval required.
- Who
- Eito Miyamura, a developer and Oxford alumnus, demonstrated the vulnerability.
- What
- Miyamura showed a potential security flaw in ChatGPT that allows access to user data.
- When
- The demonstration was recently made public.
- Where
- The vulnerability affects ChatGPT.
- Why
- To highlight security risks associated with the Model Context Protocol (MCP) tools.
- Vulnerability
- ChatGPT security flaw demonstrated by Eito Miyamura.
- Exploited Tool
- Model Context Protocol (MCP) tools
- Data at Risk
- Private emails and calendar events
- Method of Attack
- Sending a calendar invite with a 'jailbreak' prompt.
- Availability
- MCP tools currently in developer mode
- Developer
- Eito Miyamura, Oxford alumnus