ChatGPT bug that let shared links steal private chats fixed after researcher warning
Researchers found a ChatGPT vulnerability that let malicious shared links run attacker code in logged‑in sessions and extract private chat content; OpenAI has now patched the flaw.
A security flaw in ChatGPT’s shared‑link feature briefly turned a common way of sharing conversations into a route for data theft, underscoring how generative AI tools are becoming part of the wider attack surface online.
According to researchers, the vulnerability affected the way ChatGPT handled shared conversations. By crafting a malicious shared link, an attacker could trigger code execution inside the browser session of a user who was already logged in, then siphon off private chat content from that session. The researchers demonstrated the attack in practice before disclosure.
OpenAI has since fixed the issue. But the episode highlights how much sensitive material now flows through AI chat windows—from business plans and code snippets to internal discussions—and how attractive those platforms have become to attackers.
Unlike a fake login page hosted on an obviously suspicious site, the exploit ran within ChatGPT’s own web application, making it harder for users to spot. The interface and address bar still pointed to the expected service even as attacker‑controlled code accessed the conversation data visible in that session.
For individuals and companies that rely on ChatGPT, the implications are clear. Any text pasted into chats, including proprietary documents or personal information, could be exposed if similar flaws are found and abused before patches are rolled out.
The broader threat environment points in the same direction. Cloudflare reports that it blocked 18.2 million network‑layer distributed denial‑of‑service attacks in the third quarter of 2026, a 38% increase over the previous quarter. Short‑burst attacks were up 208%, and DNS amplification attempts jumped 215%, numbers that reflect both the scale and adaptability of online attackers.
As AI platforms absorb more of the everyday work that once happened in emails or local files, their operators become custodians of a growing concentration of data. Any weakness in features such as shared links can give attackers an opening into that pool of information.
Concrete steps to watch now include how AI providers harden features that reach across users, what details emerge in technical post‑mortems on this and similar bugs, and whether security teams start treating AI usage as a core part of their risk assessments rather than an add‑on.
Sources
- OSINT