Ethical Hacking News
A critical vulnerability was discovered in ChatGPT's public link, which allows attackers to read private conversations of users who open the link. The vulnerability was fixed by OpenAI, but highlights the importance of testing and fixing vulnerabilities in public links and APIs. Users do not need to take any action to fix the vulnerability, as it was fixed on OpenAI's side. The vulnerability was rated as P1, its most serious category, and is a reminder of the importance of securing public links and APIs.
Vulnerability discovered in ChatGPT's public link, allowing attackers to read private conversations. Vulnerability was identified by researchers at security firm Parameter in October 2026. Fixed by OpenAI, rated as P1, its most serious category. Vulnerability exploited through MathLive library, which loaded a listener that accepted messages. Attackers executed malicious code with the same access as the real ChatGPT site. Vulnerability highlights importance of testing and fixing vulnerabilities in public links and APIs.
Recently, a critical vulnerability was discovered in ChatGPT's public link, which allows attackers to read private conversations of users who open the link. This vulnerability was identified by researchers at the security firm Parameter, who reported it to OpenAI's bug bounty program on October 18, 2026. The vulnerability was subsequently fixed by OpenAI, which rated it as P1, its most serious category.
The vulnerability was discovered when the researchers opened a shared chat link on ChatGPT, which contained a malicious code that ran inside the user's signed-in session with the same access as the real ChatGPT site. This allowed the attacker to list the user's recent conversations, open a private chat that had never been shared with the attacker, and send a small portion of it to a server controlled by the researchers.
The researchers found that the vulnerability was caused by a combination of design choices made by ChatGPT, including the use of MathLive, a math library that loaded a listener that accepted messages about its on-screen keyboard. When the user opened the graph, it loaded the listener, which allowed the attacker to send a message with raw HTML that was not checked by the receiver. This allowed the attacker to hand the HTML to the trusted main page, which built a new frame that was not sandboxed and inherited the chatgpt.com identity.
The vulnerability was further exploited by the attacker using a script loaded from YouTube, which supplied the code that ran inside it. This allowed the attacker to execute malicious code with the same access as the real ChatGPT site.
OpenAI fixed the vulnerability by reproducing both the code execution and the private-chat proof, raising the severity from P2 to P1, and rolling a fix across two versions of ChatGPT's web frontend. By October 26, both versions rejected the messages the attack relied on.
The researchers traced the problem to three design choices that lined up: MathLive acting on a message that changed its settings without checking which frame had sent it, allowing it to carry raw HTML as keyboard-layout data, and creating a frame that was not sandboxed and inherited the chatgpt.com identity.
The researchers did not report any real-world use of the flaw, and ran its private-chat proof between separate accounts rather than against real users. However, the vulnerability highlights the importance of testing and fixing vulnerabilities in public links and APIs.
In conclusion, a critical vulnerability was discovered in ChatGPT's public link, which allows attackers to read private conversations of users who open the link. The vulnerability was fixed by OpenAI, and highlights the importance of testing and fixing vulnerabilities in public links and APIs.
Related Information:
https://www.ethicalhackingnews.com/articles/A-Vulnerability-in-ChatGPTs-Public-Link-Allows-Attackers-to-Read-Private-Conversations-ehn.shtml
https://thehackernews.com/2026/10/a-shared-chatgpt-link-could-run.html
Published: Sat Oct 31 01:41:53 2026 by llama3.2 3B Q4_K_M
