For years, "Zoom is encrypted" and "Zoom cannot see your video" were not the same sentence. Transport encryption to Zoom's servers is real and necessary. It still leaves a world where the meeting infrastructure can access plaintext media if it must mix, record, or transcribe in the cloud. After a very public period of scrutiny, Zoom shipped an end-to-end encryption mode for meetings where the keys that protect audio and video are generated and distributed among participants, not retained as a standing capability of the mixer. That is a product trade-off disguised as a checkbox: some cloud features cannot exist if the cloud cannot read the stream.

Who generates the key

In Zoom's E2EE design, a meeting leader's client plays a central role in key generation. Participants receive the meeting key through a path that is not "the server invents it and hands it to everyone." Clients can display a security code so humans can compare that they have the same key, which is an old Signal-style idea applied to a many-party meeting. The server still authenticates users and routes ciphertext. It should not be able to decrypt RTP just by being on the path.

Group meetings make this messier than a two-party chat. People join late, leave, and rejoin. A new participant needs the current key. A departed participant should not keep decrypting forever, which implies rotation. Rotation mid-meeting is a latency and glitch problem: everyone must switch keys without a drop that users will blame on Wi-Fi. Leader crash is a leadership election problem wearing a padlock.

Features that cannot come along for the ride

Cloud recording, phone dial-in, some live transcription, and join-before-host behaviors collide with E2EE. Zoom's honest engineering move was to disable or alter those features in E2EE meetings rather than pretend a cloud recorder can exist without a key. Enterprises that require recording for compliance cannot flip E2EE and keep the same workflow. That is not a documentation footnote; it is the architecture.

There is also a trust boundary at the client. E2EE does not save you from a compromised endpoint or from a participant who is allowed in. It saves you from a class of server-side access. Identity of participants still depends on Zoom's account system unless you add your own comparison of security codes.

Failure modes of meeting E2EE

The concrete failure is a host who enables E2EE, invites a PSTN user, and discovers the meeting cannot satisfy both constraints. Another is a client that falls back to transport-only encryption on an old version while the UI still says "encrypted," which is how you get a mixed-mode meeting with a false sense of the threat model. Mid-size steal: make the security level visible and block mixed modes, or clearly mark them.

Operational gotcha: key display codes that nobody checks are theater. If you are using E2EE because of a real adversary, teach people to compare codes out of band. If you are not, maybe you wanted TLS and access control, not E2EE. Leader key custody on a laptop that sleeps will stall joins; define what happens when the leader is gone. Logging that includes enough metadata to reconstruct attendance is still a privacy surface even when media is opaque. Do not log joining IP plus exact timestamps into an open warehouse and call the meeting E2EE as if that were the whole story. Test recording-disabled behavior with your legal team before a regulated customer enables the flag in a meeting that "always records."

What you can borrow

  • Separate "encrypted in transit to our servers" from "servers cannot decrypt," and name them differently in the product.
  • Disable cloud features that require plaintext instead of smuggling a server key into "E2EE."
  • Rotate group keys on membership change, and design the glitch, not just the cryptography.
  • Surface a comparable security code; without it, users cannot detect a mediator.