The facts
Parameter demonstrated the issue between separate controlled accounts and workspaces. Its tests showed code running in the ChatGPT browser session and reading an unshared private conversation. [1][2]
The disclosed chain combined the shared-conversation preview and an embedded keyboard's message handling. This was browser-session code execution, not demonstrated operating-system or OpenAI-server compromise. [1][2]
The researchers' timeline says the issue was reported October 18; OpenAI reproduced it and raised it to P1 on October 21, then confirmed both web frontends fixed on October 26. [1]
The fix was on the web service; the checked reports do not identify a client software update or a public CVE number. [1][2]
Relevant context
The October 28 disclosure is covered after the initial news window because its privacy impact remains important. The later report does not turn it into a new October 31 unpatched incident. [1][2]
Still unconfirmed
- No independently documented real-world exploitation or victim count was established in the checked sources.
Sources
References support the numbered facts above. Source checks: 1 Nov 2026, 00:57 UTC.
The actual researcher disclosure and independently owned The Hacker News report were opened. The latter attributes the technical evidence to Parameter and is not an independent reproduction. Fix confirmation and P1 classification are attributed to the researchers' account of OpenAI's response; no separate public OpenAI incident notice was retrieved.
- We found a gap at OpenAI
Parameter / Mukul Goyal · Primary source · 28 Oct 2026
- A shared ChatGPT link could run code and expose private chats
The Hacker News · Reporting · 31 Oct 2026