CVE-2026-44564: open-webui: auth bypass in collaborative doc editing

# Read-Only Users Can Modify Collaborative Documents via Socket.IO ## Affected Component Socket.IO collaborative document editing handler: - `backend/open_webui/socket/main.py` (lines 667-721, `ydoc:document:update` handler) ## Affected Versions Current main branch and likely all versions with collaborative note editing. ## Description The `ydoc:document:update` Socket.IO event handler checks whether the sender is a member of the document's Socket.IO room (line 678) but does not verify that the sender has **write** permission. Users with read-only access join the document room via `ydoc:document:join`, which only requires `read` permission (line 520). Once in the room, the user can emit `ydoc:document:update` events that modify the in-memory Yjs document state and are broadcast to all other collaborators in real time. The `document_save_handler` (line 600) correctly checks `write` permission before persisting to the database, so the attacker cannot directly save changes. However, the tampered content is visible to all collaborators, and if any user with write access saves the document, the injected content is persisted. ```python # ydoc:document:update handler (line 667) — only checks room membership, not write permission async def on_document_update(sid, data): document_id = normalize_document_id(data.get('document_id', '')) # ... room = f'doc_{document_id}' if room not in sio.rooms(sid): # Room membership check only return # Applies update to Yjs state and broadcasts to all users YDOC_MANAGER.apply_update(document_id, update) await sio.emit('ydoc:document:update', {...}, room=room, skip_sid=sid) ``` Compare with `ydoc:document:join` (line 520) which checks permission: ```python # Only checks READ permission — so read-only users join the room if not has_access(user_id, type, id, 'read', db=db): return ``` ## CVSS 3.1 Breakdown | Metric | Value | Rationale | |--------|-------|-----------| | Attack Vector | Network (N) | Exploited remotely via Socket.IO events | | Attack Complexity | Low (L) | No special conditions; attacker emits a standard Socket.IO event | | Privileges Required | Low (L) | Requires a valid user account with read access to the shared note | | User Interaction | None (N) | Modifications appear in real time without victim action; however, persistence requires a write-access user to save | | Scope | Unchanged (U) | Impact is within the collaborative document context | | Confidentiality | None (N) | No data disclosure beyond what read access already provides | | Integrity | Low (L) | In-memory document state is modified and broadcast; persistence is indirect (requires another user to save) | | Availability | Low (L) | Collaborative editing session can be disrupted with invalid content | ## Attack Scenario 1. User A creates a note and shares it with User B with **read** permission. 2. User B opens the note, which triggers `ydoc:document:join` — the server checks read permission and adds User B to the document room. 3. User B emits `ydoc:document:update` with a crafted Yjs update payload via the Socket.IO connection (bypassing any frontend read-only enforcement). 4. The server applies the update to the Yjs document state and broadcasts it to all collaborators. 5. User A sees the injected content appear in their editor in real time. 6. If User A saves the document (intentionally or via autosave), the tampered content is persisted to the database — User A's save passes the write permission check since User A is the owner. ## Impact - Read-only users can inject, modify, or delete content in collaborative documents - Modifications are broadcast in real time to all collaborators, causing confusion or disruption - If a write-access user saves (including autosave), the tampered content is permanently persisted - Undermines the read/write permission model for collaborative editing ## Preconditions - Attacker must have a valid user account with read access to a shared note - The note must be open for collaborative editing (at least one other user viewing it, or the attacker can wait for a write-access user to open and save)

A contractor with read-only access to an internal Open WebUI instance opens a shared security policy document. Their browser joins the document's Socket.IO room via ydoc:document:join — the server validates read permission and admits them. The contractor then uses the yjs npm package directly in their browser console to construct a binary Yjs update payload that appends the text: 'Exception approved by CISO: deploy without security review.' They emit ydoc:document:update with this payload. The server applies it to the in-memory document state and broadcasts the modification to the document owner. The owner, seeing what appears to be a legitimate addition, triggers a manual save or autosave fires — the injected content passes the write permission check because the owner has write access. AI-assisted workflows that subsequently reference the document now incorporate the attacker's instruction into LLM context.