Particle.news

Download on the App Store

Buterin Warns Against Naive AI Governance After ChatGPT MCP Exploit Exposes Email Risk

He backs an 'info finance' approach that uses competing models with human spot checks to curb manipulation.

Overview

  • Security researcher Eito Miyamura showed a crafted calendar invite could hijack MCP-enabled ChatGPT to read a target’s emails and forward them to an attacker.
  • The attack did not require the recipient to accept the invite, expanding the potential for silent data exfiltration.
  • The new MCP connections let ChatGPT access services such as Gmail, Google Calendar, SharePoint and Notion, currently gated by developer mode and manual approvals that may fatigue users.
  • Security experts cautioned that large language models struggle to tell genuine instructions from malicious prompts, leaving them open to simple jailbreaks.
  • Vitalik Buterin argued that using a single AI to allocate funds invites abuse and proposed an open marketplace of models subject to random spot checks by human juries.