A security vulnerability in ChatGPT's link sharing feature, discovered by researchers at the security firm Parameter, could have allowed attackers to execute code and access users' private chats. The flaw was reported to OpenAI on October 18 through their bug bounty program and was patched within eight days, earning a P1 severity rating, the highest level of concern.

The vulnerability involved a seemingly ordinary shared ChatGPT conversation link. Upon opening the link and clicking the Preview button, or in some cases simply opening the link, the attacker's code would run within the victim's active ChatGPT session. This code had the ability to list recent conversations and extract portions of chats, sending them to a server controlled by the researchers.

The exploit was possible due to a combination of features in ChatGPT. An HTML preview feature allowed certain codes to run in a sandboxed environment. However, an interactive graph within the shared chat loaded a math library, MathLive, which inadvertently created an opening. This library included a listener that accepted messages from any sender, allowing attacker HTML to be executed on the main page.

OpenAI's response was swift. They confirmed the vulnerability, increased its severity rating, and deployed a fix across ChatGPT's web frontend. By October 26, the vulnerability was mitigated, preventing the exploit from functioning. Users of ChatGPT are not required to take any action, as the solution was implemented server-side. OpenAI rewarded Parameter with $5,000 for their findings, which were publicly disclosed on October 28.