If your company uses an AI notetaker to record meetings, this is worth five minutes of your attention. A security researcher has disclosed that tl;dv, a widely used AI meeting assistant for Zoom, Google Meet and Microsoft Teams, left 181,874 meeting records belonging to 84,312 users accessible to any authenticated account.

Put plainly: anyone who could sign up for the service could query other organisations' meeting data. Not through a sophisticated exploit — through the product's own interface.

What was exposed

The disclosure, published by the researcher known as bobdahacker, describes meeting records being queryable across account boundaries. That matters because of what an AI notetaker actually holds. These tools do not store a bland calendar entry; they hold transcripts, summaries and action items from the conversation itself.

Think about what gets said in an average week of internal meetings: salary discussions, client negotiations, unreleased product plans, legal strategy, performance reviews, security incidents. A meeting transcript is often more sensitive than the documents a company carefully protects, precisely because people speak more freely than they write.

Why this class of tool carries unusual risk

AI notetakers occupy an awkward position in the security model of most organisations. To do their job they need deep, standing access:

  • Calendar access, so they know which meetings to join.
  • Permission to join calls as a participant, often automatically.
  • Recording and transcription rights for the full conversation.
  • Long-term storage of that transcript on the vendor's servers.

They are also frequently adopted from the bottom up. An individual employee connects one to their calendar because it is useful, and the tool inherits access to every meeting that person attends — including meetings with people who never agreed to any of it. Security teams often discover the tool exists only after something goes wrong.

What to do this week

For IT and security teams, the practical steps are straightforward:

  • Find out what is connected. Audit which AI meeting tools have calendar and conferencing permissions across your organisation.
  • Check the vendor's status. Look for a published advisory and confirm whether your account was affected.
  • Revoke what is not sanctioned. Remove integrations nobody formally approved.
  • Set a retention limit. Transcripts kept forever are a permanent liability; most conversations do not need archiving indefinitely.
  • Write an actual policy covering which tools may join meetings, and require participant consent.

For individuals: if you have connected a notetaker to a work calendar, check whether your employer permits it. Many people have quietly created a data-sharing relationship their company knows nothing about.

Background: convenience arrives before governance

This is a familiar pattern rather than a new one. A category of AI tool becomes genuinely useful, adoption races ahead of policy, and the security review happens after an incident. We saw the same dynamic with early cloud file-sharing services.

What makes AI assistants different is the concentration of sensitive material. A notetaker vendor accumulates transcripts from thousands of companies in one place — an unusually attractive target. Meanwhile the industry is moving toward giving assistants more autonomy, not less: the agent model depends on tools holding standing permissions and acting without a human approving each step.

Why it matters

Most AI security conversation focuses on exotic risks — models leaking training data, prompt injection, misuse of frontier capabilities. This incident is a reminder that the ordinary failures are still the ones that bite. Access control on a web application is a solved problem in principle, and it still went wrong here.

The lesson generalises: every AI tool you connect is a new place your data lives, governed by someone else's engineering standards. That is a business decision, not merely a productivity one — and the same logic behind adopting phishing-resistant sign-in applies to third-party integrations.

Key takeaways

  • A researcher disclosed that AI notetaker tl;dv exposed 181,874 meeting records across 84,312 users to any authenticated user.
  • Meeting transcripts often contain more sensitive material than formal documents.
  • AI notetakers require calendar access, call-joining rights and long-term transcript storage — a wide permission footprint.
  • These tools are usually adopted individually, so security teams may not know they are in use.
  • Audit connected tools, set retention limits, and require consent before a bot joins a meeting.

The bottom line

AI notetakers are genuinely useful, and this is not an argument for banning them. It is an argument for treating them like what they are: a third party sitting silently in your most candid conversations, storing every word. That deserves the same scrutiny as any other vendor holding your confidential data.