An AI notetaker bug let strangers join 181,000 live meetings uninvited. Some of your project calls run the same kind of bot
A security researcher found that tl;dv, an AI meeting assistant used by more than 80,000 accounts, exposed live join links for 181,874 meetings — letting anyone walk into an active call uninvited. The same class of bot now sits in on OAC meetings, GMP negotiations, and claims calls.
A security researcher found that tl;dv, an AI meeting assistant with more than 84,000 user accounts, left the join links to live video calls exposed to any other user of the platform — letting a stranger walk into an active meeting uninvited. If your project team lets an AI bot sit in on OAC meetings, GMP negotiations, or subcontractor coordination calls, this is the failure mode you're trusting the vendor to have solved.
What exactly did the flaw expose?
Researcher BobDaHacker reported that tl;dv's Firebase/Firestore backend had access controls on most collections — users, transcripts, recordings, billing — but not on the "meetings" collection. That gap exposed metadata for 181,874 completed and in-progress meetings across 84,312 users and 35,003 email domains, including .gov domains from governments in 23 countries. At any given moment, roughly 1,000 of those records showed a live call actively recording — with a conference ID an outsider could grab and use to join.
What was and wasn't reachable:
- Exposed: meeting and conference IDs (the join links for Google Meet or Microsoft Teams calls), participant email addresses and domains, and meeting status.
- Not exposed: recordings, transcripts, AI-generated notes, passwords, or billing data, according to the researcher's writeup.
To demonstrate the impact, the researcher pulled a live conference ID from the exposed data and joined an in-progress Google Meet run by Malaysia's Ministry of Education, with over 150 people on the call, unannounced.
Why does this matter for a GC's meeting habits?
Most project teams have gotten comfortable with a bot in the room. Otter, Fireflies, Fathom, Zoom's AI Companion, tl;dv — one of these is probably already auto-joining your OAC meetings, precon calls, and subcontractor coordination sessions to spit out minutes and action items nobody has to type. That comfort is exactly the exposure. A bot's presence is already normal in most calls, so an extra silent attendee wouldn't register as unusual to anyone on the call.
The calls with real cost if that goes wrong aren't the daily standups. They're the ones where numbers and legal exposure get said out loud:
- GMP negotiations and buyout calls, where pricing gets discussed before it's in a contract.
- Bid-day and bid-leveling calls between an estimator and a sub.
- Change order and claims discussions, which can become discovery material later.
- Owner calls where schedule slippage or performance issues get discussed candidly before they're written up formally.
None of those calls need an AI bot with an unverified security posture sitting in.
What to actually check before letting a notetaker into a call
This isn't a reason to ban AI meeting tools — it's a reason to stop treating them as interchangeable. Before a notetaker bot gets calendar access to sensitive project calls, someone should be able to answer:
- Does the vendor isolate meeting metadata between customer accounts, not just recordings and transcripts? Ask directly — this is the exact control tl;dv had missing.
- Has the vendor had a security disclosure in the past year, and how long did it take them to fix it? tl;dv's researcher says the report sat for about six months before a fix landed.
- Who on your team decided which calls the bot auto-joins, and is there a standing exclusion list for negotiation, claims, and legal calls?
- If the bot were compromised, what's actually recoverable by an attacker — join links, transcripts, or both? Get the vendor's answer in writing.
The AI notetaker isn't going away — it saves real time on minutes and follow-ups. But "it takes good notes" and "it's safe to have in every room" are two different claims, and only one of them has been tested here.
Related: a security firm scanned 25,000 AI-agent connectors like the ones now linking Procore and Revit and found exploitable holes in nearly three-quarters of them — the access problem isn't unique to one vendor's product.
Forward this to the person on your team who auto-approves every calendar invite for the meeting bot.
Construction AI Brief publishes three times a week. Subscribe at constructionaibrief.com.
- What happened with the tl;dv AI notetaker security flaw?
- A researcher found that tl;dv's backend database had no tenant isolation on its meetings collection, exposing metadata — meeting IDs, live join links, and participant emails — for 181,874 meetings across 84,312 users and 35,003 email domains. Anyone with a free tl;dv account could pull a live join link and enter someone else's active call.
- Were recordings or transcripts exposed in the tl;dv breach?
- No. The exposed data was limited to metadata: meeting and conference IDs, participant email addresses and domains, and meeting status. Recordings, transcripts, AI-generated notes, passwords, and billing data were not accessible through this flaw, according to the researcher's report.
- Is the tl;dv vulnerability fixed?
- tl;dv says it deployed a fix and had it validated by an independent penetration-testing vendor. The researcher's account says the flaw was first reported in January 2026 and remained unpatched for roughly six months before a fix went in around the time of public disclosure in August 2026.
- Should construction firms stop using AI meeting notetakers?
- Not categorically, but firms should stop treating notetaker bots as universally safe to invite. The practical move is to keep them out of GMP negotiations, bid-day calls, and claims or legal discussions, and to ask any notetaker vendor how meeting metadata — not just recordings — is isolated between customer accounts.
- Does this affect other AI meeting assistants like Otter or Fireflies?
- This specific flaw has been documented for tl;dv, not for other vendors. But the underlying bug class — a multi-tenant SaaS database missing per-customer access rules — is common across the category, which is why the fix is a vendor question to ask, not a reason to single out one product.