Every category of service gets breached eventually. What makes this one different is what is in the database.
Why a breach here is worse
A breached retailer leaks your address and the last four digits of a card. Unpleasant and manageable.
A breached AI companion app leaks what you typed to something you believed was private, attached to the email address you used to sign up. Often that is sexual content. Sometimes it concerns real people you named. Frequently it is the most candid material about your own life that exists anywhere in writing.
That is not a privacy inconvenience. It is the raw material for blackmail, and it cannot be reissued the way a card can.
What has actually happened
Muah AI, 2024
The significant one. A breach exposed a large volume of user email addresses alongside chat prompts. Reporting at the time noted that the exposed prompts included requests for illegal material, which produced a second problem on top of the first for anyone in that dataset.
There was no visible, meaningful security response afterwards, and the service does not clearly identify an operating company — so there is no named entity to hold responsible and no obvious regulator with jurisdiction.
This is why Muah AI scores 2.5 out of 10 on our privacy criterion, the lowest score anywhere on this site, and why we have deliberately placed no link to it anywhere, despite the commission we would earn.
Replika and the Italian regulator
A different kind of failure. In 2023 Italy’s data protection authority, the Garante, ordered Replika to stop processing Italian users’ data, citing concerns around age verification and the legal basis for processing. It subsequently issued a fine of several million euros against the operator, Luka, Inc.
Not a breach — a regulatory finding about how data was handled. It is on Replika’s record, and it is also why Replika’s documentation and deletion process are now among the most mature in this category. Regulatory action tends to produce exactly that effect.
Replika still scores 8.0 on privacy, second highest here. A company that was forced to fix something and did is in better shape than one that has never been examined.
The pattern worth noticing
Both cases show the same thing: what protects you is not the privacy policy. It is whether anyone with power can compel the company to act.
Muah AI had a privacy policy. It also had no identifiable operator and no regulator able to reach it. Replika had a regulator, got fined, and improved.
If you have used an affected service
In order of urgency.
- Change the password on that account, and anywhere you reused it. Start with your email, because that is the account that resets everything else. If you reused it, assume the reuse is known.
- Check haveibeenpwned.com. Free. Tells you which breaches your address appears in and sets up alerts for future ones.
- Turn on two-factor authentication on your email and anything financial.
- Expect blackmail emails, and do not engage. Sextortion messages following breaches of this kind are routine. The overwhelming majority are bluffing — recycled addresses from old leaks and a threatening template. Do not pay. Do not reply. Report to Action Fraud on 0300 123 2040 or at actionfraud.police.uk.
- Make a written erasure request under UK GDPR. Our data rights guide has a template. It will not undo a breach, but it limits what is there for the next one.
- Report to the ICO at ico.org.uk if the company did not notify affected users. Organisations in scope are required to notify the ICO of qualifying breaches within 72 hours, and to tell affected individuals where the risk to them is high.
How to reduce the damage in advance
Nothing here is sophisticated. All of it would have substantially limited the harm from the Muah AI breach for most of the people in it.
- A dedicated email address used for nothing else. This single step turns a breach from a personal crisis into a disposable account.
- A unique password. Assume the database leaks eventually.
- A virtual card from Monzo, Revolut, Starling or similar.
- No real names — not yours, and particularly not other people’s.
- Prefer apps that name an operator. It is the one factor that determines whether anything can be done afterwards.
How we score this
Privacy and UK jurisdiction carries 20% of our overall score, which is high for this category and deliberate.
The scale runs on one question: could you, realistically, hold someone responsible?
- 8+ — named operator in a jurisdiction you can enforce against, clear policy, working deletion.
- 6–7 — named operator, but overseas, so enforcement is awkward.
- Below 6 — no identifiable operator, or an unresolved incident.
- Below 3 — a documented breach of intimate user data with no adequate response.
Ten of the sixteen apps reviewed do not clearly identify an operating company, including Candy AI, FantasyGF, EVA AI and Muah AI. You can see every sub-score on the methodology page, or sort the whole field by this criterion alone on our privacy ranking.