Real AI chat leaks: what got exposed and how to keep your name out of the next one
Is ChatGPT safe for personal information? Usually your chats stay private. But in the last few years, real conversations have ended up in search results, public feeds and leaked databases. Here’s what happened, from the original reports.
ChatGPT shared links showed up in Google search
July–August 2025
ChatGPT’s Share button had a “Make this chat discoverable” checkbox. Fast Company found thousands of shared conversations in Google results, some with very personal details about relationships, health and work. The chats didn’t carry the user’s account name, but their contents could identify people. OpenAI called the feature a “short-lived experiment”, removed it on August 1, 2025, and worked to remove indexed chats from search engines.
Sources: Ars Technica, Aug 2025 · Search Engine Land, Aug 1, 2025 (OpenAI statement)
What Hushtag would have changed: In a shared chat that had been hushed, your name, phone, email and address would have appeared as tags like #PERSON_1 and #PHONE_1, so a stranger reading it would get the story but not who you are. It wouldn’t have stopped the leak itself.
Grok share links: more than 370,000 chats in Google
August 2025
Pressing Share in xAI’s Grok created a public page that search engines could index, without a warning to the user. Forbes counted more than 370,000 indexed conversations, including questions about medicine and psychology, and in some cases a user’s name, personal details, a password and uploaded files. Google’s Bard had a similar problem with shared chats; Google removed them from search in 2023.
Sources: Forbes, Aug 20, 2025 · BBC, Aug 21, 2025
What Hushtag would have changed: Hushtag works on Grok, so pasted text would have been hushed: your name, phone, email and address would have appeared as tags like #PERSON_1 and #PHONE_1, so a stranger reading it would get the story but not who you are. It wouldn’t have stopped the leak itself. Uploaded files are a separate step: hush them in the web app first.
Meta AI’s Discover feed: private questions posted publicly
June 2025
The Meta AI app has a public Discover feed. Reporters found people who appeared to have posted personal conversations there without meaning to, including medical and legal questions, home addresses and phone numbers, sometimes tied to their Instagram profile. Chats are private by default and sharing takes several steps. Meta later added an extra warning before posting.
Sources: WIRED, June 2025 · BBC, June 13, 2025 · Business Insider, June 16, 2025
What Hushtag would have changed: The extension doesn’t run in the Meta AI app, but you can hush text in the web app and paste the result. Then a post shared by mistake shows tags, not your details. Your profile name and photo would still show.
OmniGPT: a reported breach of 34 million chat lines
February 2025
A hacker posted what they said was OmniGPT’s data on a hacking forum: more than 34 million lines of user conversations, about 30,000 user emails and phone numbers, and links to uploaded files. OmniGPT is an “aggregator” that gives access to several AI models in one place. The company hadn’t commented when this was reported.
Sources: SiliconANGLE, Feb 12, 2025
What Hushtag would have changed: Your account email would still be in the stolen data. But in conversations you’d hushed first (with the web app on sites the extension doesn’t cover), the people and numbers you wrote about would be tags.
DeepSeek left a database of chat logs open to the internet
January 2025
Security firm Wiz found a DeepSeek database reachable with no password. It held more than a million lines of logs, including chat history in plain text, secret keys and backend details. DeepSeek locked it down shortly after Wiz contacted the company. Wiz said it couldn’t rule out that others found it first.
Sources: Wiz Research, Jan 29, 2025 · Reuters, Jan 29, 2025
What Hushtag would have changed: Hushtag works on DeepSeek, so in plain-text logs your name, phone, email and address would have appeared as tags like #PERSON_1 and #PHONE_1, so a stranger reading it would get the story but not who you are. It wouldn’t have stopped the leak itself.
A court order kept “deleted” ChatGPT chats
May–September 2025
In the New York Times’ copyright lawsuit, a US court ordered OpenAI on May 13, 2025 to preserve ChatGPT output logs that would otherwise have been deleted, including chats users deleted. OpenAI says the order ended on September 26, 2025, and that it still securely stores a limited set of April–September 2025 data for the case. A deleted chat isn’t always gone right away.
Sources: OpenAI, June 5, 2025 (updated Oct 22, 2025) · Court order, May 13, 2025 (PDF)
What Hushtag would have changed: A kept copy of a hushed chat holds tags like #PERSON_1, not the real details. The key to restore them stays in your browser, not with the AI company.
A ChatGPT bug showed people other users’ chat titles
March 2023
A bug in an open-source library let some ChatGPT users see titles from other active users’ chat history, and possibly the first message of a new chat. OpenAI said the same bug may have shown some Plus subscribers’ names, email addresses and some billing details to other users during a nine-hour window. It took ChatGPT offline and fixed the bug.
Sources: OpenAI, March 24, 2023
What Hushtag would have changed: A chat title or first message built from hushed text would have shown tags instead of names. Account details like your email are outside what Hushtag can protect.
How to keep your name out of the next one
- Treat Share as Publish. A share link can be opened by anyone who gets it, and some have ended up in search engines. In ChatGPT you can review and delete old links under Settings › Data controls › Shared links.
- Check where a chat app posts. Some apps, like Meta AI, have public feeds. Look before you tap Share or Post.
- Deleting isn’t instant. Companies keep deleted chats for a while, and a court can order them kept longer. See how to stop AI training on your chats.
- Leave the identifying details out. The AI rarely needs your name, phone, email or address to help. Here’s a real example, run through Hushtag’s detector:
You paste:
What the AI sees:
If any of the leaks above had included this chat, a reader would see a heating complaint from #PERSON_1, not you.