Open WebUI: Unauthenticated WebSocket Access to Collaborative Document Handlers (ydoc:awareness:update, ydoc:document:leave)
3.1
Low
CVSS 3.1 base
What to do
If you run open-webui 0.6.16 to before 0.10.0, upgrade to 0.10.0 or later.
Anyone who can reach it over the network, with a low-privilege account and no user interaction, given conditions outside the attacker's control, can change some data.Read from the CVSS 3.1 vector
To exploit it
What it gets
My analysis of this finding: root cause, the vulnerable code, the impact and the fix, in the site’s format. The official advisory of record is linked above.
| Package | Ecosystem | Vulnerable | Fixed in |
|---|---|---|---|
open-webui | pip | >= 0.6.16, < 0.10.0 | 0.10.0 |
Open WebUI’s real time collaboration runs on a Socket.IO server that accepts every connection, authenticated or not, and two of its Yjs document handlers never check who is calling. An unauthenticated attacker who knows or guesses a document id can inject awareness data into a live collaborative note room and spoof presence and departure events for arbitrary users. The connection itself is never gated, so there is no credential anywhere in the path.
The server is created with always_connect=True, and the connect handler in backend/open_webui/socket/main.py only decorates a session with user info when a valid token happens to be present. It never rejects a tokenless connection:
@sio.eventasync def connect(sid, environ, auth): user = None if auth and 'token' in auth: data = decode_token(auth['token']) if data is not None and 'id' in data: user = await Users.get_user_by_id(data['id']) if user: SESSION_POOL[sid] = {...} # only populated on success; connection still acceptedThe bugtokenless connections skip this and are still accepted
The two document handlers then trust whatever they are sent. ydoc:awareness:update rebroadcasts attacker controlled document_id, user_id and update into the room with no SESSION_POOL lookup and no room membership check:
@sio.on('ydoc:awareness:update')async def yjs_awareness_update(sid, data): document_id = data['document_id'] user_id = data.get('user_id', sid) update = data['update'] await sio.emit('ydoc:awareness:update', {'document_id': document_id, 'user_id': user_id, 'update': update}, room=f'doc_{document_id}', skip_sid=sid)The buguser id is taken from the client payloadThe bugrebroadcasts into the room with no membership check
ydoc:document:leave is the same shape: it takes a client supplied user_id, removes the user from the document, and emits a ydoc:user:left event into doc_{document_id}, again with no authentication and no check that the caller belongs to that room.
An attacker opens a WebSocket with no token and targets any active document room. Note rooms key on doc_{document_id}, and note ids follow the predictable note:{uuid} shape, so an attacker who learns or guesses a shared note’s id is in.
Presence spoofing and impersonation. Because user_id is taken straight from the payload, the attacker broadcasts awareness updates (cursors, selections) attributed to any user they name, and fires ydoc:user:left events for real collaborators, disrupting and confusing an active editing session.
Sink into the frontend. The awareness update blob is reflected verbatim to every other client in the room. If the collaborating frontend renders any of that attacker controlled awareness data unsafely, this unauthenticated broadcast becomes a delivery channel into other users’ browsers.
Fail closed at the boundary. The connect handler should reject any connection without a valid token instead of silently accepting it, and every ydoc:* handler should verify the caller against SESSION_POOL and confirm actual membership of doc_{document_id} before acting on or rebroadcasting the payload. user_id must be taken from the authenticated session, never from client supplied data, so presence and leave events cannot be forged.
Sanaan Fayaz Wani (GitHub sfwani) reported this vulnerability to the open-webui maintainers under coordinated disclosure and is credited as a reporter in GHSA-gmfw-g93r-vg53, published 2026-07-02.
All published findings: advisory index.