cd /news/ai-safety/chat-ask-ai-leak-300m-messages-from-… · home › topics › ai-safety › article
[ARTICLE · art-171748] src=shattered.io ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

Chat & Ask AI Leak: 300M Messages From 25M Users

An exposed database tied to the Chat & Ask AI app left roughly 300 million private chatbot messages from more than 25 million users readable by anyone who found it, according to an independent researcher identified as Harry, as summarized by the Business and Human Rights Resource Centre. Harry said he extracted and analyzed a sample covering 60,000 users and 1 million messages, and the app, which routes users to OpenAI's ChatGPT, Anthropic's Claude and Google's Gemini, claimed more than 50 million users, meaning the affected base is about half that figure by the two published numbers. Coverage from Cybersecurity Insiders, MacRumors and WIBC 93.1 FM placed the failure at the third-party app rather than the model makers' core systems, which were not reported as directly breached.

recommendation

Treat wrapper apps as part of your data supply chain: audit third-party chatbot apps' database exposure and access controls before sending them user conversations, since this one left roughly 300 million messages readable.

more in playbook: Securing agents and AI tools
by read20 min views2 publishedNov 18, 2026
Chat & Ask AI Leak: 300M Messages From 25M Users
Image: Shattered (auto-discovered)

An exposed database tied to the Chat & Ask AI app left roughly 300 million private chatbot messages from more than 25 million users open to anyone who found it, according to an independent researcher who goes by Harry. The case is described in a Business and Human Rights Resource Centre summary of the reported AI app data breach. The app connects users to large language models including OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Gemini. It was not reported as a direct breach of those three companies’ core systems.

That distinction matters for anyone who saw headlines suggesting Gemini, Claude and ChatGPT themselves fell. Based on the reports available on November 18, 2026, they did not. Coverage from Cybersecurity Insiders, MacRumors and WIBC 93.1 FM places the failure at a third-party app that sits between users and the models. Still, the people who typed into that app lost control of what they wrote. This analysis separates what is confirmed from what is not, then looks at what the leak means for ordinary users, security teams and the wrapper-app market.

The leak also fits a pattern we have tracked in recent weeks. Our report on two Fireworks AI breaches in 18 days that reached Notion chat transcripts showed how vendors further down the AI stack can expose data that users assumed was private. Chat & Ask AI adds a consumer-facing example with far larger raw numbers.

Don't miss new tech stories on Google

Add Shattered once in the Google app and our stories show up in your news suggestions.

Add now

What Researchers Say Was Exposed in the Chat & Ask AI Leak #

Harry, the independent security researcher behind the find, said he identified an exposed database associated with Chat & Ask AI. By his account, the database held approximately 300 million messages tied to more than 25 million users. Those messages were reportedly users’ private conversations with the app’s chatbot, which means the content ran from casual questions to whatever people chose to confide in an AI assistant they believed was a closed channel.

The researcher did not stop at counting records. He said he extracted and analyzed a sample covering 60,000 users and 1 million messages. That sample is the reason the story carries weight. A bare claim of a large exposed database is easy to dispute, but a pulled and analyzed slice shows the data was readable by an outsider and that the totals were not pure guesswork.

Chat & Ask AI itself claimed more than 50 million users. If the researcher’s figure of more than 25 million affected users holds up, the database touched about half of that claimed base. We calculated that ratio from the two published figures, so treat it as a rough scale marker rather than a reported statistic. No outlet in our review gave a breakdown of how many affected accounts were in the United States versus elsewhere.

Cybersecurity Insiders, MacRumors and WIBC 93.1 FM each framed the event as a user-privacy problem rather than a hack of the model makers. We follow the same framing here. The facts that carry the story are the researcher’s identification of the database, the volume he reported, and the sample he analyzed.

The Numbers Behind the Exposure #

Raw totals tell only part of the story, so the table below lays out each reported figure next to its source and what it measures. The last two rows hold our own arithmetic, clearly marked, to give the numbers some human scale.

Metric Figure Who reported it What it measures
Messages in exposed database About 300 million Researcher Harry Private user chats with the app’s chatbot
Users in exposed database More than 25 million Researcher Harry People whose chats were stored
Users in analyzed sample 60,000 Researcher Harry Slice he extracted and reviewed
Messages in analyzed sample 1 million Researcher Harry Messages inside that slice
Claimed total user base More than 50 million Chat & Ask AI Company’s own user claim
Average messages per user (calculated) About 12 Our arithmetic 300 million divided by 25 million
Exposed share of claimed users (calculated) About 50% Our arithmetic 25 million divided by 50 million

Twelve messages per user on average sounds modest, but averages hide the tail. Heavy users of chat assistants often hold hundreds of exchanges, while one-time visitors leave a single prompt. The sample ratio points the same way, since 1 million messages across 60,000 users works out to roughly 17 messages per sampled user. Neither figure tells us how deep any single person’s exposure went.

ChatGPT, Claude and Gemini: What Is Confirmed and What Is Not #

The framing that Google Gemini, Anthropic Claude and ChatGPT suffered a major breach does not match the reporting. What the reports describe is a separate app that routes user prompts to those models. The app’s database, not the model makers’ systems, was the reported point of exposure. The table below keeps the line between confirmed facts and open questions in one place.

Claim Status Basis
Chat & Ask AI had an exposed database Reported by researcher Harry’s identification, carried by multiple outlets
Database held about 300 million messages from more than 25 million users Reported by researcher Harry’s account
Researcher analyzed 60,000 users and 1 million messages Reported by researcher Harry’s account
App claims more than 50 million users Company claim Chat & Ask AI’s own statement
Exposed material was users’ private chatbot messages Reported Coverage of the researcher’s findings
Gemini, Claude or ChatGPT suffered a major breach Unconfirmed No such report in the sources
Millions of direct Gemini, Claude or ChatGPT users were victims Unconfirmed The exposure involved the app’s users
Financial loss, identity theft or account takeover resulted Unconfirmed No harm documented
Google, Anthropic or OpenAI confirmed a breach of their systems Unconfirmed No such statement found

Readers should hold two thoughts at once. The model makers appear untouched, and the users of this particular app are the ones with the problem. A person who chatted with Claude through Chat & Ask AI is in a different position from a person who logged into Claude directly, even though the underlying model might be the same.

How an AI Wrapper App Becomes a Data Risk #

A wrapper app takes a model from a major lab, adds an interface and often a subscription, and ships the result under its own brand. That business is legitimate and popular. It also creates a second custody point for user data. When you type into a wrapper, your words reach the wrapper’s servers first, then travel to the model provider. If the wrapper stores history, as most chat apps do, that archive becomes a target in its own right.

Some reports describe the root cause as a misconfigured Google Firebase database that lacked proper access controls. Treat that as a reported detail, since the sources we reviewed do not include a full postmortem from the developer. A misconfiguration of that kind is a familiar failure mode. A cloud database ships with permissive settings during development, nobody tightens them before launch, and the data sits reachable by anyone who knows where to look.

The model labs spend heavily on securing their own platforms, and they publish policies on data retention and training use. A small app developer may spend little on either. Users rarely see that gap, because the chat window looks identical whether the product is built by a lab with thousands of engineers or a team of a few. The brand name of the underlying model can even create false comfort. Seeing a famous model name in the app, people assume the famous company’s protections travel with it. They do not.

Where the Data Sits at Each Step

Follow one prompt through a wrapper. It leaves the user’s phone and lands on the app’s backend. The backend may log it, tie it to a user ID and a timestamp, then forward it to the model provider through an API. The reply comes back the same way and gets stored beside the question. At every hop, someone holds a copy. In the Chat & Ask AI case, the reported exposure sat at the app’s own storage, which is the hop users can see least.

Why Chat Histories Are More Sensitive Than Passwords #

A leaked password can be reset in minutes. A leaked conversation cannot be unsaid. People use AI assistants to draft emails about workplace disputes, ask about symptoms, rehearse difficult talks, debug proprietary code and think through money problems. Few of them treat a chat window as a public space. That habit of candor makes stored chat logs unusually revealing when they spill.

Even without names attached, conversation text can identify a person. A prompt that mentions an employer, a city, a family member’s name or a rare medical situation narrows the field quickly. Combine that with timestamps and any account details held alongside the chats, and re-identification becomes a puzzle rather than a leap. The sheer volume in this case, roughly 300 million messages, gives anyone with access plenty of material to cross-reference.

We have no confirmed report of harm from this exposure. No outlet documented identity theft, account takeover or financial loss tied to it. Absence of reported harm is not proof of safety, though. Exposure and exploitation are separate events, and the gap between them can run for months. Privacy damage from leaked chats often arrives as embarrassment, blackmail attempts or targeted phishing that references real details, and those harms are hard to trace back to one source.

Who Is Affected and Who Is Not #

The affected group, per the researcher’s figures, is the set of more than 25 million users whose messages sat in the exposed database. If you never installed Chat & Ask AI, nothing in the reports suggests your ChatGPT, Claude or Gemini history was touched. People who use those assistants through the official apps or websites are outside the described exposure. That is the cleanest answer to the question many readers are asking today.

Gray areas remain. Some reports say the developer behind this app ships other products and that similar problems could have touched those sibling apps. We could not confirm that from the verified material, so we do not treat it as established. Anyone who installed AI chat apps from small or unfamiliar publishers should assume the same open question applies until the developer publishes a clear statement.

Businesses face a quieter version of the risk. Employees often use consumer AI apps on personal phones to summarize meetings or rewrite documents. If a worker pasted client data into a wrapper that later leaked, the employer inherits a disclosure problem without ever having approved the tool. That shadow-IT angle is the part security teams should worry about most, because it sits outside every approved vendor list.

Historical Context: AI App Exposures Since 2025 #

Chat & Ask AI is not the first AI app to leave user data exposed, and the research trail is growing. Barrack AI published a tally of AI app data breaches since January 2025, titled to cover 20 incidents with the same root causes. The title alone makes the point. Repeated failures with shared causes suggest a structural problem, not a run of bad luck.

Trackers such as the AI Failure Index’s data leakage category and the Vibe Graveyard data breach tag catalog similar events, which lets readers compare this case against earlier ones. We link them as reference collections and have not audited every entry.

On this site, the pattern shows up in several forms. We reported that an Uber breach hit 149,346 drivers through Fireworks AI, and that a La Poste Maileva hack exposed data on 22 million people. Different victims, different attackers, same lesson. Data moves through chains of vendors, and the weakest link sets the real security level for everyone upstream.

What sets the current case apart is the content. Earlier incidents mostly exposed identity records, license images or business data. A leaked chatbot archive exposes thought, not just identity. People tell assistants things they would not put in a search box. That shift raises the stakes of every future exposure of this kind.

Market Impact: Trust, Procurement and App Store Pressure #

No public financial data ties this exposure to a price move at Google, Anthropic or OpenAI, and we found no official statement from any of the three. The market effect shows up elsewhere, in trust. Wrapper apps sell convenience and sometimes lower prices, but their pitch depends on users believing their chats are safe. One high-profile database exposure can undercut that pitch across the whole category, even for developers who did nothing wrong.

Procurement teams are the likeliest to react first. A company that allows AI tools at work will ask harder questions about where prompts are stored, who can read them and how long they stay. Vendors that can answer with specifics, such as retention windows, encryption at rest and independent audits, gain ground. Vendors that answer with marketing language lose it.

App store operators face quieter pressure. Every major mobile store hosts hundreds of AI chat apps, many of them wrappers. Reviewers can check for obvious policy violations, but they cannot audit backend configuration at scale. Cases like this one will feed calls for security disclosures in app listings, similar to the privacy labels already required for data collection.

The model makers sit in an awkward spot. Their brands appear in the app’s pitch, yet they did not control the failed system. Expect them to stress official channels and first-party privacy controls in coming communications. We have not seen such statements tied to this case, and we are not predicting their wording.

Competitive Comparison: Wrapper Apps vs Direct Provider Apps #

Users choose between going straight to a model maker and using a third-party app that bundles access. The comparison below describes the structural differences in custody and accountability. It is a general framework, not a claim about any specific company’s practices, and it makes no assertion about the security of any individual product.

Factor Direct provider app Third-party wrapper app
Copies of your chat Held by the model maker Held by the wrapper and passed to the model maker
Number of companies handling data One At least two
Published privacy and retention policy Usually detailed and public Varies widely by developer
Who controls the database settings The model maker’s security team The wrapper developer
Brand accountability High, tied to a major name Lower, tied to a small brand
Typical appeal Latest features first Bundled models, simple pricing
Exposure path in this case Not reported as affected Reported exposed database

Direct apps are not immune to incidents, and wrappers can be run well. The point is narrower. A wrapper adds a custody step, and each added step needs its own security investment. When that investment is missing, the user has no easy way to know.

What Industry Data Says About AI Chatbot Adoption #

The scale of chatbot use explains why a single exposed archive can reach tens of millions of people. Pew Research Center, in its report on teens, social media and AI chatbots, stated: “64% of U.S. teens reported having used an AI chatbot, while 59% reported using ChatGPT and 23% reported using Gemini.” Teens are a group where privacy exposure carries particular weight, since their chats may touch on school, family and health.

Adults are using these tools too. Pew Research Center, in its report on how Americans encountered AI-related content online, stated: “13% of surveyed U.S. adults visited an AI chatbot website at least once during March 2025, rising to 20% among adults ages 18 to 29.” The same Pew report stated: “58% of surveyed U.S. adults conducted at least one search-engine query that produced an AI-generated summary of results during March 2025.”

These figures describe usage patterns from 2025 and do not measure any exposure at Chat & Ask AI. They do show how normal chatbot use has become. When a technology reaches that many people, even a small share of badly secured apps translates into millions of affected users. That math, more than any single bug, is what makes this story matter beyond one developer.

What Users Should Do Now #

If You Used Chat & Ask AI

Start by assuming your chats were readable. Delete the app’s stored history if the app allows it, and delete your account if the developer offers that option. Review what you typed. If you shared passwords, API keys, medical details or private documents, change the credentials and think through who might use the rest. Watch for phishing messages that reference details only the chats could contain, since that is the most plausible way leaked conversation text turns into harm.

For Any AI App

Treat every chat window as a place where a stranger could someday read your words. Never paste secrets, government ID numbers or client files into an assistant you cannot vet. Prefer official apps from the model makers when you handle sensitive subjects, and read the privacy page of any wrapper before you subscribe. Our guide on 10 ways to protect your money after a bank outage and hacked sites covers account hygiene habits that apply here too, such as unique passwords and alerts on financial accounts.

For more coverage of breaches, vulnerabilities and defensive practice, see our security section.

What Developers and Security Teams Should Change #

The fix for this class of failure is dull and well known. Lock down cloud databases by default. Require authentication for every read and write. Scan your own infrastructure for public buckets and open endpoints before a researcher does it for you. Log access, alert on bulk reads and set short retention windows so a future exposure holds less. A chat archive you do not keep cannot leak.

Teams building on top of model APIs also need to think about what they store and why. If the product does not need long-term history, do not keep it. If it does, encrypt it and separate identifiers from content. Our walkthrough of indirect prompt injection defense in 12 steps covers another way chat pipelines leak data, and it pairs well with storage hardening.

Enterprise defenders should inventory which AI apps employees actually use, not only which ones IT approved. Network logs and mobile device management reports usually reveal unsanctioned tools quickly. Then decide: block, sanction with a contract, or steer people to an approved alternative. Our piece on securing AI agent-to-agent communication extends the same thinking to systems where bots talk to other bots.

Five Predictions for AI App Security #

The following are our own forecasts, based on the pattern of recent incidents. They are analysis, not reported facts, and none comes from a named source.

  1. More researchers will scan AI wrapper apps. A well-publicized find invites copycats, and open databases are easy to detect. Expect further disclosures involving small AI apps over the coming months.
  2. Enterprise allow-lists will tighten. Security teams will move from permissive to approved-only policies for AI chat tools, with vendor questionnaires covering storage and retention.
  3. App stores will face disclosure demands. Advocates will push for AI-specific security and retention labels in listings, extending the privacy label model.
  4. Wrapper apps will compete on privacy claims. Some will advertise no-logging modes. Some of those claims will later prove false, which will trigger more scrutiny.
  5. Regulators will cite chat leaks in AI privacy debates. Lawmakers in the United States and Europe already discuss AI oversight, and a case with this many users gives them a concrete example.

What We Still Do Not Know #

Several questions remain open, and honest coverage has to say so. We do not know whether anyone other than the researcher accessed the database. We do not know how long it sat exposed. We have no confirmed count of how many of the 25 million users are in the United States or any other country. And we have no statement from Google, Anthropic or OpenAI about the case, so we cannot say how those companies view their relationship with the app.

Reports suggest the developer moved to close the hole after being alerted, but the sources we could verify do not give a full account of the fix or a notification plan for affected users. Whether the company will tell users directly, and what it will say, is the next thing to watch. Regulators in several regions require breach notices in certain cases, though whether this event meets any legal threshold is not something the reports address.

The takeaway: the headline risk is real but narrower than the viral version. Gemini, Claude and ChatGPT were not reported as breached. A popular app that wraps them reportedly left about 300 million messages from more than 25 million users exposed. For readers, the practical rule is simple. Your privacy is only as strong as the weakest company that stores your chat.

Frequently Asked Questions #

Did ChatGPT, Claude or Gemini suffer a data breach?

No report in our review says so. The exposure involved Chat & Ask AI, an app that connects users to those models. It was not reported as a direct breach of OpenAI, Anthropic or Google systems, and none of the three has confirmed such an event.

How many users and messages were exposed?

Researcher Harry said the database held approximately 300 million messages from more than 25 million users. The app claims more than 50 million users overall. Harry said he analyzed a sample of 60,000 users and 1 million messages.

What kind of data was in the database?

The material reportedly consisted of users’ private messages with the app’s chatbot. Beyond that, we avoid listing specific content types, since the verified reporting does not itemize them.

Has anyone been harmed by the exposure?

No specific financial loss, identity theft or account takeover has been confirmed. That does not rule out future harm, so affected users should watch for targeted phishing and other follow-on abuse.

Is my regular ChatGPT, Claude or Gemini account at risk?

Nothing in the reports links this exposure to accounts created directly with the model makers. If you only use the official apps and sites, you fall outside the described incident. Good habits still apply, including strong passwords and two-factor sign-in.

What should I do if I used Chat & Ask AI?

Delete your history and account where possible, change any passwords or keys you pasted into chats, and stay alert to messages that quote private details. Check the developer’s official channels for a notice about the incident.

Why do wrapper apps create extra privacy risk?

A wrapper stores or relays your prompts before they reach the model maker, which adds a second company holding your data. If that company secures its systems poorly, your chats can leak even when the model provider’s own security is strong.

── more in #ai-safety 4 stories · sorted by recency
── more on @chat & ask ai 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/chat-ask-ai-leak-300…] indexed:0 read:20min 2026-11-18 · —