Indonesian Political, Business & Finance News

Your ChatGPT Conversations Could Be Exposed to the Public, Researchers Reveal

| Source: CNBC Indonesia Translated from Indonesian | Technology
Your ChatGPT Conversations Could Be Exposed to the Public, Researchers Reveal
Image: CNBC Indonesia

Jakarta, CNBC Indonesia — A dangerous security flaw was recently discovered in the popular artificial intelligence (AI) platform ChatGPT. It could allow attackers to run malicious code in a user’s active, logged-in account simply through a fake share link.

A report released by cybersecurity company Parameter said the threat could enable attackers to access and read a victim’s private conversations without their knowledge, The Hacker News reported on Monday (2 November 2026).

Parameter’s researchers reported the discovery to OpenAI through its bug bounty programme on 18 October. OpenAI responded quickly, fixing the flaw within eight days and classifying it as P1, its most serious threat category.

The attack was carried out through an ordinary ChatGPT conversation shared using a public link. The malicious link could be distributed through email, messaging apps, support tickets or social media posts.

The attack scenario was particularly subtle. Victims did not need to install additional software, join a workspace or enter their login credentials on an attacker’s website.

In the initial version of the attack, victims only had to open the link and click the Preview button once. In subsequent tests, however, researchers found that Preview could be triggered automatically as soon as the link was opened.

Potential risks and account takeover

Once the link was opened, the attacker’s malicious code would execute commands in the victim’s active account session, with the same permissions as the official ChatGPT site. The code could scan the list of recent chats, open private conversations that had never been shared, and send some of their data to a server controlled by the attacker.

Because the code called ChatGPT’s internal APIs directly—rather than merely copying what appeared on screen—its reach was extensive.

“Essentially, we could perform any action on behalf of the victim’s ChatGPT account,” Parameter’s research team wrote in its report.

Parameter said its tests were conducted in a controlled setting. There is currently no indication that the flaw was exploited by attackers in the real world, and tests involving private conversations were carried out only between accounts belonging to the researchers’ team.

How the exploit worked

Researchers said the vulnerability originated in ChatGPT’s Preview feature, which converts HTML blocks into small web pages. The Preview page is designed to be isolated in a sandbox on a separate domain.

However, attackers could exploit another feature in the same conversation: an interactive chart that loads a mathematics library called MathLive. Loading MathLive left a listener on ChatGPT’s main page.

The vulnerability arose because the MathLive listener did not verify the origin of incoming messages and allowed raw HTML data. As a result, the Preview feature could breach the sandbox boundary and inject the attacker’s HTML into the main chatgpt.com page.

OpenAI issues a dedicated fix

OpenAI confirmed the finding and immediately raised the severity rating from P2 to P1. The company deployed a fix to two versions of ChatGPT’s web frontend to reject the messages used in the attack.

ChatGPT users were advised not to worry, as the fix had been fully applied on OpenAI’s server side. Users did not need to reinstall the app or reset their passwords.

OpenAI awarded Parameter’s researchers a US$5,000 bug bounty for discovering the critical security flaw.

More from 2 November 2026 →

Cite this article
Source: jawawa.id (https://jawawa.id/newsitem/your-chatgpt-conversations-could-be-exposed-to-the-public-researchers-reveal-1793602268)
Originally published by CNBC Indonesia (https://www.cnbcindonesia.com/tech/20261102110152-37-781452/obrolan-anda-di-chatgpt-bisa-bocor-ke-publik-peneliti-ungkap-faktanya), translated from Indonesian
Print