OpenAI Patches ChatGPT Link Bug That Let Attackers Run Code and Read Private Chats
OpenAI fixed a ChatGPT vulnerability that let malicious shared links execute attacker code inside a user’s logged-in browser session, a flaw security researchers showed could expose private chat content to anyone who weaponized the feature.
OpenAI has closed a security hole in ChatGPT that allowed attackers to weaponize shared conversation links and run their own code in a victim’s logged-in browser session.
According to security researchers, a specially crafted shared link could trigger attacker-controlled code when opened by a user already authenticated to ChatGPT. Because the code ran inside the victim’s live session, it could access private chat content and other sensitive data tied to the account.
Researchers demonstrated the attack, showing that the sharing feature could be turned into a vehicle for code injection. In practice, that meant a user who trusted a malicious link could have conversations and other session details exposed without entering extra credentials or downloading files.
For many people, ChatGPT now holds information that goes well beyond casual queries. Users often paste in draft contracts, proprietary code, internal documents, and personal material. A vulnerability that makes those chats accessible to an attacker turns a convenience feature into a channel for data theft.
The flaw underlines the security challenge facing fast-growing AI platforms. Each new sharing or integration feature expands the attack surface and requires thorough testing to prevent abuse. When a platform becomes central to work and research, a single bug in a popular feature can have wide impact.
OpenAI has patched the vulnerability, blocking the specific method of executing malicious code through shared links.
Signals to track now include any further technical detail OpenAI releases about the bug and fix, changes to its security testing or bug bounty programs, and whether other AI providers re-examine their own link-sharing and embedding features for similar weaknesses.
Sources
- OSINT