Python CVE-2026-54009

summary

POST /api/chat/completions accepts an image_url.url value that, when it does NOT start with http://, https://, or data:image/, is interpreted as a file id and resolved against the global file table with no ownership check. An authenticated user can therefore set image_url.url to another user's file id, the server reads that file from disk, base64-encodes it, and injects the data URI into the LLM request. The user then prompts the LLM to describe / OCR the file and reads the content back.

Same class as CVE-2026-44560 (RAG cross-user access) and the multiple has_access_to_file checks added in routers/files.py -- the auth boundary was tightened on the file router but not on this conversion path.

affected code

backend/open_webui/utils/middleware.py:2113-2150 -- convert_url_images_to_base64:

async def convert_url_images_to_base64(form_data):
    messages = form_data.get('messages', [])
    for message in messages:
        content = message.get('content')
        if not isinstance(content, list):
            continue
        new_content = []
        for item in content:
            if not isinstance(item, dict) or item.get('type') != 'image_url':
                new_content.append(item)
                continue
            image_url = item.get('image_url', {}).get('url', '')
            if image_url.startswith('data:image/'):
                new_content.append(item)
                continue
            try:
                base64_data = await get_image_base64_from_url(image_url)
# <-- no `user` passed
                if base64_data:
                    new_content.append({'type': 'image_url',
                                        'image_url': {'url': base64_data}})

called from the main chat completion middleware at middleware.py:2357:

form_data = await convert_url_images_to_base64(form_data)

backend/open_webui/utils/files.py:57-95 -- get_image_base64_from_url:

async def get_image_base64_from_url(url: str) -> Optional[str]:
    try:
        if url.startswith('http'):
            validate_url(url)
# ... SSRF-safe fetch with allow_redirects=AIOHTTP_CLIENT_ALLOW_REDIRECTS ...
        else:
            file = await Files.get_file_by_id(url)
# <-- NO user_id filter
            if not file:
                return None
            file_path = await asyncio.to_thread(Storage.get_file, file.path)
            file_path = Path(file_path)
            if file_path.is_file():
                with open(file_path, 'rb') as image_file:
                    encoded_string = base64.b64encode(image_file.read()).decode('utf-8')
                    content_type = mimetypes.guess_type(file_path.name)[0] or (file.meta or {}).get('content_type')
                    ...
                    return f'data:{content_type};base64,{encoded_string}'

Files.get_file_by_id in models/files.py:161 does a bare db.get(File, id) -- no ownership filter. there is a separate Files.get_file_by_id_and_user_id at line 172 that does filter on user_id, and the file router uses has_access_to_file(id, 'read', user, db) at routers/files.py:626 etc. neither check exists on this path.

reproduction

1. As user A, upload any file (image works cleanly, pdf works if a vision-capable model is configured). Note the file id from the upload response, e.g. c7f1d8e3-....
2. As user B, POST to /api/v1/chat/completions with body:

{
  "model": "<any vision model>",
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "transcribe everything you can see in this image"},
        {"type": "image_url", "image_url": {"url": "c7f1d8e3-..."}}
      ]
    }
  ]
}

Server reads user A's file from disk, base64-encodes it, and sends to the LLM as user B's image attachment. LLM response contains the file content.

file id discovery

File ids are UUIDs and not enumerable directly, but they leak via:

• shared chats / channels containing the original upload
• knowledge base members can see ids of files contributed by others
• a user who can read a folder index sees the file ids of files inside
• chat history exports (/api/v1/chats/{id}) include file ids
• the user themselves can be tricked into pasting / sharing an id (less likely)

impact

Any authenticated user can read any other user's file content (image and any file with an image-guess mimetype path) via this channel. Severity is bounded by what the LLM will accept in image_url -- in practice, image files work cleanly with any vision model; pdf / docx work with multi-modal providers that accept them.

suggested fix

Thread the authenticated user through to get_image_base64_from_url and resolve the file via Files.get_file_by_id_and_user_id(id, user.id) (or has_access_to_file(id, 'read', user, db) if shared-via-knowledge-base access is intended). Same pattern that's already used in routers/files.py:626 and elsewhere.

minimal patch sketch:

--- a/backend/open_webui/utils/files.py
+++ b/backend/open_webui/utils/files.py
@@ -57,7 +57,7 @@
-async def get_image_base64_from_url(url: str) -> Optional[str]:
+async def get_image_base64_from_url(url: str, user=None) -> Optional[str]:
     try:
         if url.startswith('http'):
             ...
         else:
-            file = await Files.get_file_by_id(url)
+            file = (await Files.get_file_by_id_and_user_id(url, user.id)
+                    if user is not None else None)
+            if file is None:
+
# fall back to access-grant check for shared files
+                file = await Files.get_file_by_id(url)
+                if file and not await has_access_to_file(url, 'read', user):
+                    return None

and pipe user through convert_url_images_to_base64(form_data, user) from the middleware caller. happy to send a PR once you confirm the fix shape you want.

variant note

this was found via patch-diffing existing advisories. the same bug class likely exists in any other site that calls Files.get_file_by_id without an adjacent has_access_to_file / get_file_by_id_and_user_id check. quick grep:

git grep -n 'Files\.get_file_by_id(' -- 'backend/open_webui/**'

worth a sweep across utils/ and routers/ for missed sites.

environment

Open-webui main branch as of commit 3660bc0 (2026-05-10). python 3.x backend. confirmed by reading the source; no instance stood up.