Are NSFW AI chats private? What the apps keep, and the Muah.ai breach

Are NSFW AI chats private? Not in the way most people assume. Your chats are stored on the company's servers, readable by the company, and on several apps used to train models or reviewed by staff when a filter flags them. In 2024 one "uncensored" app, Muah.ai, lost its database, and prompts tied to 1.9 million email addresses ended up in a hacker's hands. Extortion attempts followed.

This guide covers what happened at Muah.ai, what the apps we compare say they keep and why, and the handful of habits that limit the damage if one of them is breached next.

What happened in the Muah.ai breach?

In September 2024 a hacker took Muah.ai's database: 1.9 million email addresses linked to the prompts users had written.

404 Media broke the story on October 8, 2024¹. The hacker described the site as "a handful of open-source projects duct-taped together" and said getting into the database was easy. The data tied users' prompts, which describe their sexual fantasies in detail, to their email addresses, many of which looked like personal accounts with real names. 404 Media also found many prompts asking for child sexual abuse scenarios, which is illegal.

Have I Been Pwned, the breach-notification service run by security researcher Troy Hunt, lists the breach² as September 17, 2024, with 1,910,261 accounts and three kinds of data: AI prompts, email addresses and sexual fetishes. It flagged the breach as sensitive, so only someone who proves they own an address can check whether it is included. Hunt wrote in his weekly update³ that the child abuse material in it was "rampant".

According to Malwarebytes, summarizing 404 Media⁴, Muah's administrator blamed competitors in the "uncensored AI industry" and said moderators delete all child-related bots. Within days Malwarebytes added that the data was being used in "active extortion attempts", and the law firm Linklaters⁵ wrote about the same threats. Two years on, Muah is still running, its privacy policy⁶ is still dated May 2022, and its terms name no legal company. Our Muah AI review has the rest.

Has it happened to other AI chat apps?

Yes. In 2025 two companion apps left 43 million messages on a server anyone could open.

In October 2025 Malwarebytes reported⁷ on research by Cybernews into Chattee Chat and GiMe Chat, two apps from a Hong Kong developer. A server with no password had exposed 43 million messages and more than 600,000 images and videos from over 400,000 users, with purchase records, IP addresses and device IDs. Nobody had to hack anything.

Both cases share a pattern: small companies, thin engineering, and the most sensitive text a person can write stored next to something that identifies them. Back in February 2024, Mozilla Foundation reviewed 11 romantic AI chatbots⁸ and gave every one its "Privacy Not Included" warning.

What do NSFW AI chat apps keep about you?

Your chats, characters, account details, device data and payment records, usually for as long as your account exists and sometimes years after.

From the policies we read in November 2026:

• Kupid AI's privacy policy⁹ keeps personal data for six years after you stop being a user.
• DreamGF gives no retention period, and its terms¹⁰ take a "perpetual" and "irrevocable" license to what you create.
• Saucepan's policy¹¹ keeps chats until you delete them or your account.
• Janitor AI collects chats, characters, IP and device data per its privacy policy¹², and keeps some records, such as your email, after you delete your account.
• PolyBuzz's policy¹³ says it doesn't sell data but lists Meta, Google and AppLovin for analytics and ads, and doesn't say how long chats are kept.

Do these apps train AI on your chats?

Some say yes, one says only if you opt in, and several don't say at all.

• Opt-in only: Saucepan says training on your data is off by default and anonymized if you turn it on. That is the clearest policy in our comparison.
• Yes: Candy.ai's privacy notice¹⁴ says it may train its AI and moderation models on user interactions, with human review of de-identified chats, and that messages go to third-party LLM providers.
• With permission: PolyBuzz asks to use uploaded photos and precise location to train its AI.
• Not stated: Janitor AI, DreamGF, Kupid and Muah don't say whether chats train models.

If you use your own API key on Janitor or Saucepan, your chats also go to that model provider, under its own policy. Check whether it keeps prompts and whether it trains on them before you send anything personal.

Who can read your chats?

The company can. On most explicit apps, flagged chats can be reviewed by a person.

Moderation is the reason. Candy.ai's blocked content policy¹⁵ says flagged content and other content on the account may be reviewed manually. Kupid and DreamGF both describe human reviewers. DreamGF says staff read chats only if you report a problem; PolyBuzz's FAQ says staff can't see them. None of these apps offer end-to-end encryption, because the AI on their server has to read what you write. Our guide to what "uncensored" means explains why the filters, and the reviewers behind them, aren't going away.

How do you delete your data?

Most apps let you delete your account in settings; a few still make you email them.

• Janitor AI: delete it yourself with an email code, followed by a grace period, per its help center¹⁶. Some records, such as your email address, are kept.
• Candy.ai: delete the account in Settings, per its help center¹⁷; the privacy notice keeps account data for up to three years after your last activity.
• Saucepan and PolyBuzz: delete chats or the account in the app; PolyBuzz says backups are purged later.
• DreamGF: EU and UK users can ask for erasure, which it says it answers within 30 days.
• Muah AI: by email only, under a privacy policy written before its breach.

Deleting doesn't reach copies already taken. If your address was in a breach, deleting the account now won't remove it from the leaked data, which is why the habits below matter more than any cleanup.

How to use NSFW AI chat more privately

• Use an email address that is only for this. The Muah.ai leak hurt because prompts sat next to people's real addresses.
• Leave out real names, places, workplaces and photos, yours or anyone else's. Most apps ban real people in sexual content anyway.
• Pick a company with a name and an address. An app whose terms name no legal entity has nobody a regulator can reach.
• Check the billing name. Candy.ai bills as EverAI and DreamGF as "Dream AI"; know what will appear on a shared statement before you pay. Our cost guide covers billing and refunds.
• Delete what you don't need. Delete old chats, then the account, when you stop. Retention periods like Kupid's six years start when you leave.
• Be careful with ID checks. Upload ID only to a named age-verification vendor, never by email; see age verification for NSFW AI chat.

The safest assumption is simple: anything you type into an AI chat app could one day be read by someone else. Write accordingly.