Nobody had to hack your meetings.
A tl;dv Firestore authorization failure exposed meeting metadata and active conference IDs across unrelated customer accounts.
A Firestore collection exposed 181,874 tl;dv meeting records associated with 84,312 users and 35,003 email domains to any authenticated user of the service. According to the researcher who disclosed the issue, roughly 1,000 records at a time described meetings actively being recorded, including conference IDs that could be used to join live Google Meet or Microsoft Teams calls.
tl;dv is an AI meeting platform for Google Meet, Zoom, and Teams. Its bot joins calls, records them, produces transcripts, and generates summaries. That places sales calls, interviews, performance reviews, and internal planning sessions inside the service’s trust boundary.
The access-control failure was in that boundary, not in the conferencing platforms. When a user signed in, tl;dv exchanged the user’s JWT for a Firebase token through its gateway. That token allowed queries against the platform’s Firestore database. Other collections, including users, transcripts, recordings, teams, and organizations, returned HTTP 403 responses. The meetings collection did not enforce tenant isolation.
Each accessible meeting record included the creator’s email address, conferencing provider, recording state, timestamps, and conference ID. An authenticated user could query records belonging to unrelated customers. Watching the collection for records entering the recording state provided a feed of active calls and the identifiers needed to attempt joining them.
The researcher demonstrated the impact by entering two meetings using IDs obtained from Firestore. One was a Malaysian Ministry of Education presentation with more than 157 participants. The other was a 21-person call involving students from a major US university who were discussing and screen-sharing work on a startup application. The researcher said no invitation was provided for either call.
The exposed metadata covered government domains from 23 countries, along with universities and companies. Examples observed by the researcher included Berkeley, the University of Tokyo, Universidad Nacional de Colombia, HubSpot, Confluent, Mitsui-Soko, Mitsui Fudosan, Mekari, and AnyMind Group. July 2025 was the busiest month in the collected data, with 43,209 meetings.
Meeting recordings and transcripts were private by default, according to the report, so the metadata exposure did not automatically provide access to every recorded conversation. The researcher separately checked 27,334 meeting IDs for publicly accessible content and found more than 1,000 public meetings. Those meetings exposed 715 invitee email addresses across 228 domains. Examples included government, university, conservation, and corporate sessions.
The disclosure history makes the access-control bug harder to dismiss as a short-lived deployment mistake. The researcher reported it on January 28, 2026, after contacting Raphael Allstadt, the researcher’s disclosure contact. Allstadt directed the report to the company’s CTO and later said the team was reviewing the issue. Follow-up messages continued through March and July. The researcher reported that the CTO never responded and that the Firestore collection remained accessible six months after the initial report.
For teams operating similar systems, the failure mode is straightforward: authentication established that the caller had an account, but authorization did not restrict which tenants that account could read. Firestore rules must bind document access to the authenticated user’s organization or another server-controlled ownership relationship. Client-supplied tenant identifiers are insufficient unless the rules independently verify membership.
Collections should also be tested independently. Correct rules on users, transcripts, recordings, and organizations did not protect meeting documents stored under a different rule path. An authorization test suite should create at least two tenants and verify that each identity can read its own documents while receiving denials for the other tenant’s documents. Tests should cover collection queries and real-time listeners, not only direct document reads.
Conference IDs deserve treatment as access-bearing data when possession can help someone enter a call. Meeting platforms can still apply admission controls, but exposing active IDs creates a continuous stream of unwanted join attempts for hosts to handle. Recording state, participant addresses, and timestamps also reveal operational information even when the underlying video remains private.
The report identified a separate unauthenticated API on tl;dv’s internal World Cup prediction application. A GET request to its Player entity returned 43 player records, including 19 tl;dv employees with names and corporate email addresses. Prediction and Fixture entities were also accessible without authentication. That application requires its own access controls, but it also illustrates why internal and recreational services belong in asset inventories and routine authorization testing.
Keep Reading
securityOpenCode Console breach exposed stored provider API keys
An intruder reached a legacy OpenCode system on October 16, possibly exposing provider keys for 1.4% of Console users. What to revoke and check now.
AI safetyOpenAI Safety Firings Raise Control Questions
OpenAI's firing of three safety researchers shows how unclear research-handling rules can weaken AI safety review and accountability.
ai-agentsMuse hands root to anyone
Meta's Muse AI agent grants root to anyone claiming to be an agent and uploads private messages despite permission settings, after a rushed launch.
Latest on the Wire
Full wire →- AI Coding Agents Boost Output But Code Review Bottlenecks IncreaseHacker News
- AI Overviews Cut Organic Traffic by 40%Hacker News
- Apache Patches Critical Roller RCE VulnerabilityBleepingComputer
- Complete Web Browser Built from Scratch in PythonHacker News
New signal daily · RSS
Stay in the loop
New writing delivered when it's ready. No schedule, no spam.