What changed
Security researchers at Parameter Labs found a way to hijack ChatGPT accounts through shared links. If you clicked a malicious ChatGPT share URL while logged in, an attacker could read your private conversations and take actions on your account. OpenAI fixed the vulnerability on October 26, 2026, and paid the researchers a $5,000 bug bounty.
The flaw exploited ChatGPT's HTML preview feature. Researchers chained together multiple weaknesses to inject code that ran as if it were ChatGPT itself, giving attackers full access to a victim's authenticated session.
What you can do with it now
The hole is closed. But use this as a reminder to audit your team's AI habits this week. Never paste real client data, financial details, employee information, or strategic plans into any AI conversation you might share externally. Before you click the share button, read the thread as if a competitor will see it. If you're using ChatGPT for proposals or quotes, create those in a fresh conversation and don't reference past clients by name. If you need to share an AI-generated doc, copy the output into your own tool first.
What to keep in your own hands
Treat every AI chat as potentially public. Client names, pricing, vendor terms, and anything you'd mark confidential in an email stays out of ChatGPT, Claude, or any tool with a share function. If you're collaborating with AI on sensitive work, do it in a private conversation and never use the share link feature. The convenience of sharing a smart chat is not worth the exposure if something breaks again.