The little bot that joins your Zoom call and quietly writes the minutes has become office furniture. Most people never think about where those transcripts go. In August 2026, a security disclosure gave everyone a reason to think about it: a popular AI meeting notetaker was found to be exposing meeting records from tens of thousands of accounts — including live calls at government ministries and universities.
Here is what actually happened, how badly it affects ordinary users, and the checks worth running before your next meeting.
What happened in the tl;dv meeting leak?
A security researcher found that tl;dv, an AI notetaker for Zoom, Google Meet and Microsoft Teams, was missing a database access rule. Any signed-in tl;dv user could list meeting records belonging to other customers — 181,874 meetings from 84,312 users across 35,003 email domains, according to the researcher’s public report.
The exposure was not a hacker breaking encryption. It was a missing fence. tl;dv stores data in Google Firestore, and the researcher reported that the platform had correct access rules for users, chats, transcripts and recordings — but not for the meetings collection itself. That one gap removed the boundary between one company’s data and everyone else’s.
The records reached further than a typical startup breach. The report describes government meetings from 23 countries on .gov domains, universities including Berkeley and the University of Tokyo, and roughly a thousand meetings marked as actively recording at any given moment. The story was picked up by Dark Reading and spread quickly because of one detail: some of the calls were still in progress.
The takeaway in one line: an AI notetaker does not just record your meeting, it becomes a second copy of your company’s most sensitive conversations, sitting in someone else’s database.
How can an AI notetaker leak a live call?
Two ways, and they are very different. The first is a backend flaw — a missing tenant boundary that lets one customer read another customer’s records. The second is social: an attacker asks to be let into your call while pretending to be the notetaker bot your team already expects.
That second trick is the uncomfortable one. The researcher reported that requesting admission while impersonating an expected notetaker worked in roughly 80% of tested cases. Nobody clicked a malicious link. Somebody just saw a familiar bot name in the waiting room and hit Admit.
It works because meeting culture has trained us to wave the bot through. A participant called “Notetaker” or “Meeting Assistant” looks like infrastructure, not a guest. This is the same pattern of over-trusting automated helpers that showed up in the UK safety tests on rogue AI agents — the risk is rarely the model itself, it is the permissions we hand it without thinking.
What does tl;dv say about it?
tl;dv disputes the framing that one flaw sat unfixed for six months. In its public response, the company says a vulnerability affecting access to certain meeting metadata was identified earlier in the year through both a routine assessment by its independent penetration testing vendor, Abicom, and a responsible disclosure from an independent researcher, and that the fix was deployed and formally validated by Abicom.
The company states that the recent incident involved a previously unknown, alternative exploitation path affecting the same part of its stack — two distinct vectors, not one ignored report. The researcher’s disclosure timeline tells a less forgiving version, describing an initial report in January 2026 and follow-ups through July with no response from the CTO.
Both accounts can be partly true, and for a user the practical conclusion is the same: the security of your meeting archive depends on a vendor’s internal configuration that you cannot see, audit, or verify from the outside.
Are AI notetakers legally allowed to record you?
It depends on where the people on the call are sitting. In the United States, 11 states require all-party consent to record a conversation: California, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire, Pennsylvania and Washington. If even one participant is in one of those states, everyone generally needs to consent — not just the organiser who invited the bot.
That is not a theoretical risk. Otter.ai is defending four consolidated federal lawsuits in In re Otter.AI Privacy Litigation, where Judge Eumi K. Lee heard argument on the company’s motion to dismiss on 20 May 2026; the motion remains under submission with no settlement. Fireflies.ai faces two separate biometric privacy suits in Illinois. These cases ask a genuinely new question: whether wiretap statutes written decades ago apply to a bot that quietly joins a video call.
Illinois adds a sharper edge. Under the Biometric Information Privacy Act, a spoken disclaimer at the top of a call is not enough — extracting a voiceprint requires written notice and a written release. Elsewhere the rules differ again: the UK and EU treat recordings as personal data under GDPR, and Australia’s state surveillance-devices laws vary by state much as US laws do.
Consent is not a checkbox in the notetaker’s settings; it is something the humans on the call have to actually give.
Six checks before you let a notetaker into your next meeting
- Announce the bot out loud. Say a recording assistant is present and give people a real chance to object. A banner nobody reads is weak evidence of consent.
- Never admit an unexpected bot. If you did not schedule the notetaker, do not let it in. Treat a surprise participant in the waiting room like a surprise visitor at reception.
- Lock the waiting room for sensitive calls. Legal, HR, medical, M&A and security reviews should not have automated attendees at all.
- Check retention settings. Many tools keep recordings indefinitely by default. Set a deletion window — 30 or 90 days is plenty for most teams.
- Audit who can see the archive. Ask whether transcripts are shared workspace-wide by default, and whether departing employees lose access.
- Ask the vendor for a recent penetration test. Vendors that publish audit results and a security contact are meaningfully safer bets than ones that do not.
One more habit worth building: assume anything said on a recorded call could be read later by someone outside the room. That is not paranoia, it is what a misconfigured database rule does in practice.
Does this mean you should stop using AI notetakers?
No. For most teams the productivity gain is real, and the fix is scoping rather than abstinence. Use the notetaker for status updates, customer calls and project reviews. Keep it out of conversations covering legal advice, personnel decisions, security incidents and unannounced financials.
The broader pattern is worth noticing. As AI tools move from answering questions to acting on our behalf — joining calls, holding credentials, and in some cases operating shopping accounts — the security question shifts from “is the model accurate?” to “what does this software have access to, and who else can reach it?”
Frequently Asked Questions
Was tl;dv hacked?
Not in the traditional sense. No password cracking or malware was involved. A missing database security rule meant any authenticated user could query meeting records across other accounts, which is a misconfiguration rather than an intrusion.
How do I know if my meetings were exposed?
Individual users cannot check this themselves. Ask your workspace administrator to contact the vendor for confirmation, and treat any meeting recorded through the affected period as potentially readable by others until the vendor confirms otherwise.
Can someone join my Zoom call pretending to be a notetaker bot?
Yes, if a host admits them. The researcher found that requesting admission under a familiar notetaker name succeeded in about 80% of tested cases. Only admit bots you scheduled yourself, and lock meetings once everyone has arrived.
Is it illegal to record a meeting with an AI notetaker?
Not by default, but consent rules matter. Eleven US states require all-party consent, Illinois requires written consent before a voiceprint is collected, and GDPR treats recordings as personal data. When participants are spread across regions, follow the strictest rule that applies.
Which AI notetaker is the most secure?
No independent ranking settles this, and marketing pages are not evidence. Judge vendors on published penetration test results, configurable retention and deletion, enterprise access controls, and how quickly they respond to security researchers.
