RC RANDOM CHAOS

Researchers Chain ChatGPT Vulnerabilities into Account Hijacking

· via Hacker News

Original source

We Took over ChatGPT Accounts

Hacker News →

A research team discovered a critical vulnerability allowing attackers to hijack ChatGPT accounts through a shared conversation link. By chaining two seemingly harmless features—an interactive graph and an HTML preview—researchers bypassed ChatGPT’s sandbox isolation and content security policy to execute arbitrary JavaScript with full access to victims’ authenticated sessions. The attack required only that the victim click the malicious link, after which the payload could enumerate private conversations, retrieve sensitive content, and call ChatGPT’s APIs directly as the victim.

The exploit weaponized a series of interconnected flaws. The graph feature loaded MathLive, a math library that installed a message listener in the main ChatGPT page for keyboard events. The sandboxed preview could communicate with this listener via postMessage. However, MathLive accepted messages from any origin without validation and allowed HTML in keyboard layouts, which it then injected into the main page without sanitization. To achieve code execution despite ChatGPT’s content security policy, the researchers exploited a YouTube JSONP endpoint that the policy whitelisted, crafting a callback parameter that would execute as JavaScript.

The practical attack combined both vulnerabilities into a single shared conversation that triggered automatically when opened. OpenAI patched the issue on October 26 and rated it P1. The exploit demonstrates how multiple independently safe components can create dangerous attack chains when validation mechanisms fail across frame boundaries and dynamically generated content.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.