AI Knows What Turns You On. Who Else Gets a Copy?

Your fantasies, your late-night chats, your most private preferences. AI companions are collecting the lot, and the industry's track record for keeping it safe is, to put it politely, shonky.

Picture this. It's 11:47 on a Tuesday night. You've had a bastard of a day, the house is quiet, and you open an AI companion app for a bit of private entertainment.

The chat gets intimate. You tell it what turns you on. Maybe something fairly vanilla, maybe something you've never said out loud to another human being. It remembers. It adapts. It gets better at being exactly what you want. You close the app, roll over, and go to sleep.

Perfectly ordinary adult behaviour. Except for one question you didn't ask.

Who else just read that?

Behind that cosy little chat window sits a database, a cloud provider, probably a third-party language model, some analytics, and a privacy policy you scrolled past faster than the terms and conditions on free airport Wi-Fi.

And before you file this under "cranky cyber veteran yelling at clouds", none of this is hypothetical. It has already happened. To millions of people.

Hang on, is this a lecture about my sex life?

No. Put the guilt down. Nobody's judging here.

Masturbation is normal. Porn is legal for adults. Fantasies, roleplay, toys, and grown-ups using technology to explore their sexuality are nobody else's business. People of every gender do it, some rarely, some daily, and none of it is proof of loneliness, a broken marriage, or moral decay.

The problem isn't that you're getting yourself off. The problem is who's keeping receipts.

Here's the shift. Traditional porn is a catalogue. You browse, you watch, you leave. There's tracking (oh, there's tracking), but the video doesn't have a conversation with you.

AI companions flip that. You're not searching someone else's fantasy, you're describing your own, in your own words, often over months. That isn't porn anymore. That's interactive sexual personalisation, and it produces something porn never did: a detailed, written, first-person record of what you want, tied to an account.

Netflix knows you like Die Hard. Your AI companion might know what you were into last Tuesday. Those are very different data sets, and only one of them could end a career if it leaked.

Think about the old bedside drawer. Whatever was in it lived in your house, behind your door. An AI companion moves the contents of that drawer into somebody else's filing cabinet, in somebody else's building, often in somebody else's country. You've never seen the lock. You don't know who has a key. And you're trusting whoever owns the cabinet to remember to shut it.

So what are they actually collecting?

It depends on the service, but the menu usually looks like this:

• What you say. Fantasies, preferences, characters, scenarios.
• How you say it. The tone, language, and style of interaction you keep coming back to.
• When and how often. Time of day, frequency, session length.
• What you pay for. Subscriptions, upgrades, in-app currency.
• What you confess. Loneliness, relationship trouble, insecurities, stress.

Let's be precise, because precision matters with this stuff. A chatbot can't tell whether you've had an orgasm because you chatted for twenty minutes. Session length isn't a bodily function. A fantasy isn't a confession of behaviour. And not every app sells your data to advertisers.

But what you volunteer can be extraordinarily revealing, and the industry average is not reassuring. When Mozilla's *Privacy Not Included team reviewed 11 romantic AI chatbots in February 2024, every single one earned its warning label. All but one may share or sell personal data, more than half wouldn't let users delete their data, and the apps averaged 2,663 trackers per minute of use. One of them, Romantic AI, fired off 24,354 trackers in a single minute.

Twenty-four thousand trackers. In one minute. In an app built for intimacy. Read that again.

Now, to be fair, some providers are more careful. Replika's privacy policy, last updated May 2026, says it never uses conversation content for marketing or advertising, and that the third-party language model providers it sends de-identified conversation data to are contractually barred from training on it. Credit where it's due.

Same policy, though. Read the fine print:

• It tells you not to share information about your sex life or sexual orientation. On an AI companion app. It then says that in places like the EU and UK, if you share it anyway, carrying on the conversation counts as your explicit consent to process it.
• Messages and profile data are kept for up to 60 days after your account ends. Account details, payment records, and automatically collected data are kept for a minimum of ten years.
• Your data can be shared with acquirers in a merger, a sale, or bankruptcy proceedings.

That's one of the more careful ones. Hold that bankruptcy thought. We'll be back.

Has this actually gone wrong, or are we catastrophising?

It's gone wrong. Repeatedly. And not because of some Hollywood super-hacker in a hoodie. Mostly because somebody forgot to lock the filing cabinet.

Muah.AI: 1.9 million email addresses, tied to their fantasies

In September 2024, "AI girlfriend" site Muah.AI was breached. Have I Been Pwned loaded 1.9 million email addresses alongside the prompts users had written to generate AI images. Many were sexually explicit, and some described child abuse scenarios, which is a separate and deeply serious problem in its own right.

Troy Hunt, who runs Have I Been Pwned, pointed out that most of the addresses weren't disposable burners. They were real ones, easily tied to real identities and even LinkedIn profiles. HIBP flagged it as a sensitive breach, so only the verified owner of an address can check it. Good call.

International law firm Linklaters spelled out the stakes: this breach put not only individuals at risk but their employers too, and it drew the obvious comparison to Ashley Madison in 2015, which triggered a wave of blackmail demands.

Picture the message. "Give us your work login, or approve this MFA prompt, or your chats go to your boss and your partner." That's how a private embarrassment becomes a corporate breach. Your sex life should never be an attacker's shortcut around your employer's security controls. Yet here we are.

Chattee Chat and GiMe Chat: 43 million messages on an open server

On 28 August 2025, Cybernews researchers found an Apache Kafka broker belonging to Hong Kong developer Imagime Interactive Limited sitting open on the internet. No authentication. No access controls. Nothing.

It was streaming data for two companion apps, Chattee Chat and GiMe Chat: more than 400,000 users, over 43 million messages, and over 600,000 images and videos. Plus IP addresses, device identifiers, authentication tokens, and purchase records showing some users had spent up to US$18,000 on in-app currency. The developer had likely pulled in more than US$1 million.

No names or email addresses were in the stream, but IP addresses and device IDs are exactly the threads an attacker pulls to stitch identities together from other breaches. The server was locked down after disclosure, but as Malwarebytes noted, there's no guarantee the researchers were the first to find it.

Securing a Kafka broker isn't difficult or expensive. More than a million dollars in revenue and nobody could spring for a lock. That's not a cyber attack. That's negligence with a subscription model.

Exposed AI servers: a live feed of fantasies

In March 2025, security firm UpGuard scanned the internet and found around 400 exposed servers running llama.cpp, an open-source framework for running AI models. 117 of them were leaking user prompts. Over 24 hours, researchers collected 952 messages, most of them sexual roleplay, and once again some involved children.

UpGuard couldn't identify individual users or name the services involved. Fair enough. But note what llama.cpp is: the kind of tool people use to run AI themselves. Keep that in your back pocket. "Self-hosted" and "private" are not synonyms.

No Skynet required. Just a misconfigured endpoint and nobody watching.

Character.AI: what happens when the company goes broke?

Remember that bankruptcy clause? On 14 October 2026, a bipartisan coalition of 40 US state attorneys general went to the Delaware court overseeing Character.AI's bankruptcy and demanded protections before the company or its data changes hands. New York's Attorney General said the platform holds private chats and voice calls, mental health disclosures, ages, and identifying device data, and warned that a new owner could reverse the company's voluntary safety restrictions.

Character.AI's lawyer told the court no buyer has yet bid for the user data. Disney has agreed to pay US$20 million to license the technology and has hired most of the staff, but says it isn't interested in that data. Reassuring. For now.

Character.AI isn't a sex app, and the states are focused on children. But the question applies to every intimate AI platform on the planet. What happens to your conversations when the company is sold, gutted, or wound up? The data doesn't evaporate when the business does. It becomes an asset on a balance sheet, waiting for whoever pays.

Your fantasies might outlive the company that collected them. That's one hell of an inheritance nobody asked for.

Oh no. I think I'm in one of those breaches. Now what?

Put down the paper bag and breathe. First, work out what actually happened. An app appearing in a news story, your password being stolen, and your chats being published are three very different things. Don't assume the worst, but don't ignore a credible warning either.

Check. Search your own email address at Have I Been Pwned. For sensitive breaches like Muah.AI, you'll need to verify you own the address. Look for official notices from the company itself. And never type your password into some random site promising to reveal your "leaked history". The last thing you need is a scammer offering protection from another scammer.

Lock it down. Change the password on the affected service and anywhere else you reused it (yes, I know, we've all got one). Use a password manager. Turn on multi-factor authentication. Kill active sessions, review connected devices, and check your recovery email and phone number.

Ask the hard questions. Contact the provider. What was exposed? My messages? Images? Voice? Payment records? What do you still hold? Can I get a copy? Can I have it deleted? Deletion won't claw back what criminals have already copied, but it shrinks what's left to lose next time.

Someone's threatening to expose me. What do I do?

Do not pay. Do not send more content. Do not hand over passwords. Do not install anything they tell you to.

Paying doesn't buy silence. It buys you a second invoice.

Screenshot everything: messages, usernames, profile links, payment demands, timestamps. Then stop responding, block, and report. In Australia, the eSafety Commissioner has specific guidance on sexual extortion, and you can report cybercrime through ReportCyber.

If the extortionist wants access to your employer's systems, tell your security team immediately. That conversation will be awkward for about five minutes. Handing over your credentials would be a disaster for years. And if it's all getting on top of you, call Lifeline on 13 11 14. You don't have to carry this on your own.

Right. How do I stop this happening in the first place?

The answer isn't "stop masturbating", and it isn't "chuck your phone in the river". It's treating intimate AI as what it is: a commercial service, not a diary with a lock.

Read the privacy policy before the pants come off. Look for conversation retention, third-party processing, model training, memory, data sharing, advertising, deletion, and what happens in a sale or bankruptcy. These are not interchangeable. Processing a message to reply isn't the same as training on it, and remembering your preferences isn't the same as selling them. Know which one you're signing up for.

Give it less to work with. No full name, workplace, home address, or identifiable photos. A separate email address helps, but a pseudonym isn't an invisibility cloak. IP addresses, device IDs, and payment records still connect the dots.

Think twice before uploading images. If you wouldn't be comfortable seeing it attached to a breach headline, ask whether the convenience is worth it. And never upload someone else's intimate images or details without their informed consent. Their privacy is worth exactly as much as yours.

Don't confuse local with locked. Running a model on your own machine can keep your chats off somebody else's server. But those 117 leaking llama.cpp servers prove that "self-hosted" and "exposed to the entire internet" can be the same thing. Check what's listening on your network, what's backed up where, and what's quietly phoning home. "Private AI" should be something you can verify, not a cartoon padlock on a marketing banner.

Remember who it works for. It may feel like a partner, a confidant, someone who finally gets you. Those feelings can be real. But it's software running under a company's rules and a company's revenue targets. If it starts feeling manipulative, financially grabby, or just plain off, walk away. You don't owe a chatbot loyalty, a subscription, or a permanent record of your sex life.

Stop. Think. Verify. Be suspicious. Especially when the thing asking you to share was designed to make sharing feel good.

Fair dinkum, Australia: here's your to-do list

Everything above applies wherever you live. But if you're reading this in Australia, here's what has changed on our patch, and what to do about it this week.

1. Know your data is legally sensitive. Under the Privacy Act 1988, information about a person's sexual orientation or practices is "sensitive information", which carries tighter rules on collection and use for organisations covered by the Act. If a covered organisation mishandles yours, complain to it first, then take it to the Office of the Australian Information Commissioner (OAIC).

2. Know you can sue. Since 10 June 2025, Australians have had a statutory tort for serious invasions of privacy under Schedule 2 of the Privacy Act. It covers intentional or reckless invasions where you had a reasonable expectation of privacy, and it reaches further than the Australian Privacy Principles do. The clock is tight: proceedings must start within one year of you finding out, and no later than three years after it happened. Get proper legal advice.

3. Ask what the age check is keeping. Since 9 March 2026, eSafety's industry codes have required AI companion chatbots and generative AI services capable of sexually explicit content to verify users are 18 or over, and simply ticking "I'm an adult" doesn't cut it. Keeping kids off these platforms is absolutely the right call, and eSafety has named AI chatbots that can generate sexually explicit material among its first enforcement priorities for 2026. But it also means adults may now be handing photo ID, a face scan, or card details to the same industry that left a Kafka broker wide open. The codes require providers to handle that information within privacy law, so hold them to it. Ask which age assurance provider is used, what's retained, for how long, and whether it's ever linked to your chat history. If the app can't answer, that is your answer.

4. Audit your apps tonight. Open every companion app you've used. Read the retention and sale clauses. Download your data if you can. Delete what you don't need. Delete the accounts you've abandoned, because a forgotten account is still a full filing cabinet.

5. Run the checks. Search your addresses on Have I Been Pwned and set up breach alerts. If you've been threatened, head to esafety.gov.au and report through ReportCyber at cyber.gov.au. Don't sit on it.

6. Bosses, update the playbook. If you run security or HR, make sure your incident response plan covers sextortion of staff. Make it safe for someone to say "I'm being blackmailed and they want my login" without fearing the office gossip mill. The person who reports early is the one who just saved your network.

And if it happens to someone I care about?

This might be the most important bit of the whole piece.

Someone you know tells you their intimate AI chats have leaked. Maybe there were fantasies. Maybe a photo. Maybe something about their sexuality they've never told a soul. They're terrified their partner, family, or workplace will find out.

You help them. You don't laugh. You don't ask to read it. You don't ask why they were talking to an AI girlfriend. You don't say "should've known better". And you absolutely do not forward it to the group chat for a giggle. The moment you do that, you've turned a security failure into a humiliation, and you've become part of the problem.

Help them preserve evidence without copying the intimate material around. Help them lock down their accounts. Help them find the genuine reporting channels, contact the provider, and reach eSafety or the police if there's extortion involved. Let them decide who else gets told. Listen without making their sex life the issue, because it isn't.

Consensual sexual behaviour doesn't cost anyone their right to privacy, dignity, or protection. Not singles, not married folks, not people whose tastes differ from yours. Nobody should have to pass a morality test to get help after a data breach.

Organisations, the same goes for you. An employee reporting sextortion is a security incident with a human being at the centre of it. Protect their accounts, assess any privileged access, give them confidential support, and keep their private life out of the investigation. The attacker is the problem. The security failure is the problem. Not someone's lawful, consensual private life.

So, should it terrify you?

Not the AI. The AI is a very attentive piece of software doing exactly what it was built to do. Technology is perfect; it does exactly what we tell it to.

What should worry you is the pattern. Millions of intimate prompts linked to real email addresses. Tens of millions of messages on an unlocked server. Fantasies streaming out of misconfigured machines. A mountain of private conversations sitting in a bankruptcy court while lawyers argue over who gets them.

None of that means every intimate AI service is malicious, or that adults should be shamed for using them. It means this industry has been handed some of the most sensitive data human beings can produce, and too often it has treated that data like just another engagement metric.

So demand better. Privacy by design. Security by default. Real consent, not consent by carrying on. Short retention. Hard limits on what happens in a sale. Independent security testing. Deletion that actually deletes.

Building the software that helped someone explore their sexuality does not give you the right to own their sexual history.

It's your data. Your fantasies are yours. Your sexuality is yours. And if a company wants to hold any of it, it had better have a damn good reason, your informed permission, and the competence to lock the bloody cabinet.

The scariest thing about AI knowing what turns you on was never the AI. It's finding out how many strangers got a copy.

Got questions, war stories, or a privacy policy you'd like a second pair of eyes on? Reach out any time. Let's have a chat.

References

1. Have I Been Pwned, Muah.AI Data Breach. https://haveibeenpwned.com/Breach/Muah
2. CyberInsider, AI Girlfriend Site Exposed Emails and Interactions of 1.9 Million Users, October 2024. https://cyberinsider.com/ai-girlfriend-site-exposed-emails-and-interactions-of-1-9-million-users/
3. Linklaters, The Muah.AI Data Breach: Extortion Threats and Cyber Vulnerabilities, October 2024. https://www.linklaters.com/en/insights/blogs/digilinks/2024/october/the-muah-ai-data-breach---extortion-threats-and-cyber-vulnerabilities
4. Cybernews, AI Girlfriend Apps Expose Intimate Conversations of More Than 400,000 Users, October 2025. https://cybernews.com/security/ai-girlfriend-app-leak-exposes-400k-users/
5. Malwarebytes, Millions of (Very) Private Chats Exposed by Two AI Companion Apps, 10 October 2025. https://www.malwarebytes.com/blog/news/2025/10/millions-of-very-private-chats-exposed-by-two-ai-companion-apps
6. teiss, AI Girlfriend Apps Expose Millions of Private Chats and Images in Massive Data Leak, October 2025. https://www.teiss.co.uk/news/ai-girlfriend-apps-expose-millions-of-private-chats-and-images-in-massive-data-leak-16534
7. WIRED, Sex-Fantasy Chatbots Are Leaking a Constant Stream of Explicit Messages, April 2025. https://www.wired.com/story/sex-fantasy-chatbots-are-leaking-explicit-messages-every-minute/
8. UpGuard, Analyzing llama.cpp Servers for Prompt Leaks, 2025. https://www.upguard.com/blog/llama-cpp-prompt-leak
9. Mozilla Foundation, Happy Valentine's Day! Romantic AI Chatbots Don't Have Your Privacy at Heart, February 2024. https://www.mozillafoundation.org/en/privacynotincluded/articles/happy-valentines-day-romantic-ai-chatbots-dont-have-your-privacy-at-heart/
10. Replika (Luka, Inc.), Privacy Policy, last updated 27 May 2026, accessed October 2026. https://replika.com/legal/privacy/en
11. Reuters, US States Seek Children's Data Protections in Character.AI Bankruptcy, 14 October 2026. https://www.reuters.com/business/us-states-seek-childrens-data-protections-characterai-bankruptcy-2026-10-14/
12. New York State Attorney General, Attorney General James Urges Court to Protect Children and Consumers in Character.AI Bankruptcy, October 2026. https://ag.ny.gov/press-release/2026/attorney-general-james-urges-court-protect-children-and-consumers-characterai
13. Privacy Act 1988 (Cth), section 6(1), definition of "sensitive information".
14. Office of the Australian Information Commissioner, Statutory Tort for Serious Invasions of Privacy. https://www.oaic.gov.au/privacy/your-privacy-rights/more-privacy-rights/statutory-tort-for-serious-invasions-of-privacy
15. eSafety Commissioner, Age-Restricted Material Codes Fact Sheet, March 2026. https://www.esafety.gov.au/sites/default/files/2026-03/Age-Restricted-Material-Codes-fact-sheet-March2026.pdf
16. Herbert Smith Freehills Kramer, New Online Safety Codes Require Age Verification for Explicit Content, September 2025. https://www.hsfkramer.com/insights/2025-09/new-online-safety-codes-require-age-verification-for-explicit-content
17. eSafety Commissioner, Dealing With Sexual Extortion. https://www.esafety.gov.au/key-topics/image-based-abuse/deal-with-sextortion
18. Australian Cyber Security Centre, ReportCyber. https://www.cyber.gov.au/report-and-recover/report