AUTO-UPDATED

We Took over ChatGPT Accounts

Researchers discovered a critical vulnerability in ChatGPT that allowed attackers to hijack user accounts and steal private conversation data simply by sending a malicious public link.

Key Points

  • The exploit bypassed ChatGPT’s sandbox by chaining together the "Preview" feature, a math library called MathLive, and an unsandboxed iframe.
  • Attackers could execute arbitrary JavaScript within a victim's authenticated session without requiring the user to install software or open developer tools.
  • The vulnerability allowed unauthorized access to private conversation history and the ability to perform actions on behalf of the victim's account.
  • OpenAI patched the security flaw on October 26, 2026, after classifying the report as a P1 critical issue and awarding a $5,000 bounty.
  • The attack demonstrated how seemingly harmless features, when combined, can create a path that circumvents Content Security Policy (CSP) protections.

Why it Matters

This incident highlights the significant risks inherent in complex web applications where multiple independent features interact across security boundaries. It serves as a reminder that even robust sandboxing can fail if individual components trust messages from unverified sources, potentially exposing sensitive user data to unauthorized third parties.
Parameter.ai Published by Mukul Goyal
Read original