Patched ChatGPT Bug Let Malicious Shared Links Steal Private Chats from Logged‑In Users
OpenAI fixed a ChatGPT vulnerability that allowed attacker code in malicious shared links to run in a user’s logged‑in session and access private chat content, according to researchers who demonstrated the exploit.
A recently fixed flaw in ChatGPT shows how quickly AI chat services are becoming targets for familiar web attacks, with one bug briefly allowing malicious shared links to tap into users’ private conversations.
Security researchers found that a crafted shared chat link could execute attacker-controlled code inside a victim’s active ChatGPT session. Because the exploit ran in the context of a logged-in user, it could expose chat histories and other data visible in the browser, according to a technical report.
The problem stemmed from how ChatGPT handled content in or loaded from shared conversations. If a user opened a malicious shared link while signed in, injected scripts could run with the same privileges as the user’s session. This kind of behavior is a known category of web application vulnerability, and its appearance in a widely used AI interface underlines how existing attack techniques are being turned on new platforms.
For individual users, the risk was that sensitive information pasted into ChatGPT — such as work projects, legal drafts, or personal details — could be captured by someone who convinced them to click a hostile link. Many people treat AI chats as private workspaces, but the flaw showed that the underlying web session can be a weak point if not properly secured.
Organizations that use ChatGPT for support, coding help, or document drafting face similar exposure. Staff often share proprietary or regulated material in chats. A compromise of even one user account through a malicious link could provide an attacker with select, high-value snippets of internal information without breaching core corporate systems.
The incident adds to evidence that AI platforms are now important targets in their own right because of the kinds of data they handle and how quickly they’ve spread through workplaces. As happened with content management systems and cloud dashboards, attackers are probing AI interfaces for configuration errors and unsafe sharing features.
Researchers publicly demonstrated the attack, and OpenAI has since patched the flaw. The case highlights the role of independent security testing in keeping rapidly deployed AI tools in line with long-established web security standards.
Security teams will be watching to see whether OpenAI releases more detail on the vulnerability, how regulators fold AI-specific risks into existing data-protection rules, and whether companies begin isolating AI use into dedicated browsers or accounts to limit damage if similar bugs appear in the future.
Sources
- OSINT