How personal data becomes exposed

"Data leak" covers very different events. A colleague forwarding a link, a software bug and a stolen password have different causes and different fixes. Telling them apart shows you which precautions help with which risk.

You share it, and it travels further than you meant

The most ordinary route involves no attacker. In ChatGPT, anyone with the link to a conversation shared from a personal account can view it, and deleting the link doesn't remove copies recipients have saved. Managed workspaces handle access differently. Whether a link can be found by a search engine is a separate question, and a shared link isn't automatically public to everyone. It only takes one forward, though. OpenAI on shared conversations

Search can still become part of the story. In July 2026, Reddit users found chats people had shared from Claude listed in Google results, some with names and addresses. Anthropic said conversations are private by default and that it doesn't give search engines directories of shared chats. The search trick stopped working, but links already out there stayed live. Fortune on shared Claude chats

What helps: read a conversation before you share it, and remove old links.

A third party receives it

Tools can pass your text on. OpenAI says data sent through a GPT action falls under the recipient's policy and may be kept longer than the chat itself. The person using the tool may never notice that a second organisation is involved. OpenAI temporary chat

A browser extension can do it too. In December 2025, OX Security reported two Chrome extensions, with over 900,000 users between them, that copied a real AI sidebar and sent people's ChatGPT and DeepSeek chats to the attackers' servers. They had asked permission to collect "anonymous, non-identifiable analytics data". OX Security on the extensions

What helps: find out who else receives the data before you connect an app or action, or install an extension.

A bug shows it to someone else

A security researcher found that changing a number in a Meta AI request could return another user's private prompt and answer. Meta fixed the bug in January 2025 and said it found no evidence of abuse. TechCrunch on the Meta AI bug

Bugs can also carry data past a rule meant to stop it. In early 2026, Microsoft 365 Copilot summarised emails with confidentiality labels from users' Sent Items and Drafts. Microsoft said this gave nobody access to information they weren't already allowed to see. Nobody did anything wrong on the user side in either case. BleepingComputer on the Copilot bug

What helps: anything you never sent can't appear this way. Sharing less limits what a failure like this can touch.

Someone logs in as you

The 23andMe case wasn't about AI, and it shows what happens once an account is compromised. According to the UK's Information Commissioner's Office, reused passwords enabled a credential-stuffing attack in 2023. The investigation concerned the personal information of 155,592 UK residents, and stolen data was advertised for sale on Reddit. A person's password reused on another site ended up affecting data far beyond one account. ICO on 23andMe

Your AI account's chat history is an account like any other. Whoever logs in can read it. IBM reported that malware which steals passwords led to the exposure of more than 300,000 ChatGPT logins in 2025. IBM describes stolen logins, not a breach at OpenAI. IBM X-Force 2026

What helps: a unique password and a second login factor, and less history kept in the account. Replacing IDs in your text doesn't secure your login.

A model repeats what it learned

In tests published in January 2026, researchers got production language models to reproduce long passages of books they had been trained on, some only after getting around their safeguards. That work tested books, not personal data. An older study on GPT-2 did pull out personal details that came from public sources. Together they show the mechanism exists. They don't show that a prompt you type today is trained into a model straight away, or that every service is equally vulnerable. Ahmed et al., 2026, Carlini et al., USENIX Security 2021

What helps: know your service's training terms, and keep private details out of what it could learn from.

A model works it out

Inference needs no stored copy of anything. In one study, language models worked out who was behind pseudonymous online accounts from what those accounts had written. At best they matched 68% of accounts to the right profile, and 90% of their matches were correct. In another, 240 US participants struggled to anticipate what a text gave away, and their rewrites often left those risks in place. Both results apply to the models, data and tasks studied. Lermen et al., USENIX Security 2026, Wang et al., CHI 2026

This is also why replacing names isn't the end of the job. The European Commission explains that details which identify someone when combined are personal data. European Commission on personal data

What helps: read for context. A school, a diagnosis and a small town can identify someone with no name in sight.

What happens next

Exposure rarely ends with the first person who sees the data. It can be copied, combined with other records, sold or used to make a scam message convincing, and you usually can't call it back. The person affected is often someone who never used the service: a candidate, a client, a patient.

So the moment before sharing is the one you control. Later steps can only reduce the damage.

Where Anonymizator fits

Anonymizator addresses one part of this: what your text and supported files contain when you share them. Replacing IDs with placeholders means less personal data for a bug, a link or a third party to expose. You review the result and choose what to share.

It doesn't secure your accounts, prevent an incident at a provider or make a text anonymous. Placeholders can be reversed with the list behind them, and what's left in the text can still identify someone. Keep that list, in Current IDs or an ID file, out of what you share.

Review before you send

AI privacy best practices