tl;dv, the AI meeting assistant, exposes 180k meetings: live spying possible

Researcher BobDaHacker has documented a critical access control failure in tl;dv (Too Long; Didn’t View), the AI assistant that records and transcribes video calls. The application’s Google Cloud Firestore backend leaves the “meetings” collection without tenant isolation: any authenticated user can query the metadata of every meeting where tl;dv is active. BobDaHacker then demonstrated that with those metadata, an attacker can join a call by impersonating the bot itself. The test succeeded in 80% of attempts. The findings were made public on 4 August 2026. The vendor has yet to respond, and the vulnerability remains live.

The Firestore database without rules: an open door

tl;dv uses Firebase as its backend to store meeting information. The “meetings” collection holds identifiers, timestamps, and access data for every video call where the bot is present or scheduled. Normally, Firestore security rules should restrict access to users within an organisation, isolating tenants. In this case, no rules were applied: any authenticated user can enumerate the collection without restrictions, reading every meeting’s metadata. From there, gaining access to the call is a short step.

BobDaHacker reverse-engineered the bot’s self-join mechanism and showed how an attacker can use the invitation tokens contained in the metadata to be recognised as tl;dv itself. The meeting control system accepts the fake bot and grants entry in 80% of the tested calls. The method works on Zoom, Google Meet, and Microsoft Teams – the platforms tl;dv integrates with. No complex exploit is needed; a few API calls and the missing Firestore rules are enough.

180k meetings, 80k users, 23 governments affected

Analysis of the metadata captured the scale of the problem. BobDaHacker found over 180,000 completed call records belonging to more than 80,000 users. The domains involved include .gov addresses from 23 different countries, indicating tl;dv was present in public administration meetings. The sample also includes large enterprises such as HubSpot and Mitsui Fudosan, and universities like UC Berkeley and the University of Tokyo. These are not test meetings; the metadata describes real work calls, with dates, participants, and platforms.

Transcripts and recordings are not directly exposed through this vulnerability. But real-time entry into a call allows listening to confidential conversations, harvesting proprietary information, and, in some cases, intercepting classified exchanges. The ability to systematically enumerate meetings also lets an attacker select high-value targets, focusing on government or corporate users.

No vendor response, users left alone

The researcher contacted tl;dv before public disclosure, but the company sent no response. As of 4 August 2026, the flaw remains active. The app’s public repositories show no security rule updates, and no patch has been released.

For those using tl;dv, the only immediate protection is to disable the bot’s automatic entry into calls, at least until an official fix arrives. It is also useful to review meeting logs for unauthorised participants, watching for unexpected AI assistant join times. On the vendor side, mitigation is straightforward: apply Firestore rules that restrict reads on the “meetings” collection to users from the same organisation. A measure that should have been part of the baseline configuration.