Discord DAVE flaw: encryption depended on checking who the Welcome message admitted
CVE-2026-104480 describes a libdave membership validation failure under attacker-controlled signalling. The media cipher was not broken; an unrecognised member could be admitted to the encrypted group before version 1.2.0.
By Parminder Kumar Sharma · LinkedIn · · 6 min read
An encrypted call needs the right roster
Discord’s DAVE system uses end-to-end encryption for audio and video. Its open-source libdave library had a failure at a specific trust boundary: a client could accept a Message Layer Security (MLS) Welcome that placed an unrecognised member in the encrypted group. The CVE record, published on 2 October, describes possible compromise of media confidentiality and integrity when an attacker controls the DAVE signalling path. Discord fixed the validation in libdave 1.2.0, released on 26 August.
This is not a discovery that the media cipher can be mathematically broken, or that a stranger can simply join any Discord call. The prerequisite is substantial: an attacker must control the signalling route, such as the Voice Gateway or an equivalent ability to alter the messages a client sees. If that condition meets a vulnerable client, the roster used to decide who receives encrypted media can differ from the roster the user expects. Strong encryption then protects media for the wrong group.
The public record does not establish how many client installations used the affected library, that a signalling path was compromised, or that any real call was intercepted. The report is a vulnerability disclosure, not a victim list.
Three parts of the call have different jobs
The DAVE protocol documentation separates signalling, group key management and media forwarding. The Voice Gateway distributes the messages clients use to establish or change an MLS group. MLS then manages the membership and secrets for that group. The selective forwarding unit, or SFU, relays encrypted audio and video frames without needing to read their content.
A Welcome message is how a participant receives the information needed to join an MLS group. Its content has to be validated against the participants the application recognises. The vulnerable libdave code missed that check for an unrecognised member. In the fix commit, the implementation checks the Welcome roster and rejects a mismatch. That is a membership decision. It should not be described as a generic “broken E2EE” claim without the signalling prerequisite.
The layers in the reported failure.
| Layer | Normal role | Relevant boundary |
|---|---|---|
| Voice Gateway signalling | Carries MLS setup and updates | Attacker control of this path is required |
| MLS Welcome validation | Admits members to the cryptographic group | Pre-1.2.0 could accept an unrecognised member |
| Encrypted media through SFU | Relays frames to accepted group members | The accepted roster defines the intended audience |
The control-flow mistake in one view
The source-backed path is deliberately short. An attacker with signalling control presents or alters an MLS Welcome. A vulnerable client processes it. If the resulting cryptographic group contains a member absent from the expected Discord participant list, the client should reject the transition. Before the fix, that check was insufficient. Media sent for the accepted group may then be available to the unrecognised member, and the CVE also names an integrity risk.
The visual below shows the signalling precondition, the unrecognised member in a Welcome, and the client’s before-and-after roster decisions. It does not add credential theft, server compromise, wallet malware or a remote attack from an ordinary call participant. Those would be different claims requiring different evidence. The SFU is a relay, not a secret-key holder in this account.
The disclosure date is not the patch date
The CVE was published on 2 October, but the fix was available earlier. The libdave 1.2.0 release is dated 26 August. A headline that says “Discord fixes a flaw today” would muddle disclosure with availability. The structured affected field in the CVE specifies the 1.1.0 series before 1.2.0; its prose description says versions “before 1.2.0”. Where those scopes differ, a component owner should check the actual embedded version and vendor guidance rather than assume all older packages behave identically.
Library releases and application releases are also distinct. Publishing libdave 1.2.0 does not, by itself, show when every Discord desktop, mobile or web client incorporated it. The public CVE and repository do not provide a population count or per-platform rollout table. Organisations embedding libdave should verify their own dependency and shipped artefact. End users should keep Discord clients current, while avoiding the false precision of declaring a particular device safe from the CVE record alone.
Dates and claims that should not be collapsed.
| Date or scope | Established fact | What it does not prove |
|---|---|---|
| 26 August | libdave 1.2.0 release containing the fix | Every client updated on that day |
| 2 October | CVE record published | A new exploit began that day |
| Affected field | 1.1.0 to below 1.2.0 | Exact installed version of every client |
What the Voice Privacy Code can and cannot answer
DAVE documents a Voice Privacy Code that participants can compare through a separate channel. This is a useful human check on an encrypted conversation when the users know one another and can verify the code independently. It does not replace the library’s roster validation. The CVE and patch do not show that comparing the code would always reveal or prevent this particular unrecognised-member condition.
For a developer or investigator, the decisive evidence is whether the affected library version was present, whether the attacker could control the signalling messages, and whether an unexpected MLS member was accepted. Logs and client state may be incomplete after a call, so the absence of a complaint is not a technical audit. The strongest ordinary response is to ship the fixed library and test membership transitions with deliberately mismatched Welcome data.
What a regression test should prove
The central regression test is a Welcome that otherwise parses correctly but introduces a member the application does not recognise. A fixed client should reject the group transition before trusting it for media. The test should also cover ordinary membership changes so that valid joins still work. This is a test of agreement between two views of the call: the cryptographic MLS roster and the participant list delivered by the application.
That agreement has to persist across joins, leaves and reconnections, not only at the first handshake. A client that checks one snapshot and then accepts later group changes without equivalent validation can recreate the same category of risk. This is a general engineering implication of the reported bug, not a claim that the published patch misses those cases. The fix commit is the source to inspect for the exact implementation in libdave.
Security claims should also preserve the strong attacker condition. A normal participant can influence a conversation in many social ways, but this CVE concerns control of the signalling material a client processes. Testing an ordinary invited user is not a substitute for an adversarial signalling fixture. Conversely, a simulated malicious Welcome in a laboratory does not prove Discord’s live Voice Gateway was ever under an attacker’s control.
The decision for implementers and users
Take this with you
Check the relevant layer
- Inventory the libdave version actually shipped in each application build; use 1.2.0 or later.
- Test that an MLS Welcome containing an unexpected member is rejected, including during group changes.
- Protect and monitor the Voice Gateway signalling path as a trust input, not merely a message transport.
- Keep end-user clients updated; use the Voice Privacy Code where an independent participant check is useful.
- Describe any suspected incident with evidence of signalling control and affected client state before claiming media exposure.
Sources
- PrimaryCVE-2026-104480 recordCVE Programaccessed 2026-10-02
- Primarylibdave membership validation fixDiscordaccessed 2026-10-02
- Primarylibdave 1.2.0 releaseDiscordaccessed 2026-10-02
- PrimaryDAVE protocol documentationDiscordaccessed 2026-10-02


