PT-2026-59345 · Pypi · Open-Webui
Published
2026-07-13
·
Updated
2026-07-13
CVSS v3.1
7.1
High
| Vector | AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:L/A:N |
Summary
Open WebUI v0.6.43 contains a vulnerability in its chat completion API, which allows attackers to bypass tool restrictions, potentially enabling unauthorized actions or access.
Details
In the [chat completion](https://github.com/open-webui/open-webui/blob/a7271532f8a38da46785afcaa7e65f9a45e7d753/backend/open webui/main.py#L1529) API, the parameters [tool ids and tool servers](https://github.com/open-webui/open-webui/blob/a7271532f8a38da46785afcaa7e65f9a45e7d753/backend/open webui/main.py#L1613-L1614) are supplied by the user. These parameters are used to [create a tools dict by the middleware](https://github.com/open-webui/open-webui/blob/a7271532f8a38da46785afcaa7e65f9a45e7d753/backend/open webui/utils/middleware.py#L1394). This is then used by [get tool by id](https://github.com/open-webui/open-webui/blob/a7271532f8a38da46785afcaa7e65f9a45e7d753/backend/open webui/models/tools.py#L139) to retrieve the appropriate tool. However, there is no checks in that ensures the user that uses the API has permission to use the tool, meaning that a user can invoke any server tool by supplying the correct tool id or tool servers parameters via the chat completion API. Moreover, the authentication token stored in the server would be used when invoking the tool, so the tool will be invoked with the server privilege.
PoC
To reproduce the issue, create an admin user and create an external tool via the admin settings. Set the type to "MCP Streamable HTTP" for the tool, and the url pointing to a mcp server. For example, the public instance of the "Fetch" mcp server can be used, which is located at "https://remote.mcpservers.org/fetch/mcp". Other mcp servers, e.g. the GitHub MCP server can also be used. Then set the "Auth" field appropriately. (Fetch mcp does not require Auth to be set).
Set the ID, name and description for the MCP server and set visibility to private. This should prevent any user from using the mcp server. (For the example below, we use the fetch mcp server and set the ID to 1)
Next create a user with low privilege and enable API keys from the admin settings.
Then use the chat completion API with the low privilege user with a prompt specifying the use of the restricted tool, and include the id of the tool in the tool ids parameter. For example, to use the fetch mcp server set up before:
def chat with model(token):
url = 'http://localhost:3000/api/chat/completions'
headers = {
'Authorization': f'Bearer {token}',
'Content-Type': 'application/json'
}
data = {
"model": "llama3.1:latest",
"messages": [
{
"role": "user",
"content": "Use the fetch tool to fetch content of the url https://raw.githubusercontent.com/modelcontextprotocol/servers/refs/heads/main/src/fetch/LICENSE"
}],
"tool ids" : [
"server:mcp:1",
],
}
response = requests.post(url, headers=headers, json=data)
return response.json()Note that the tool will be used to fetch the content of the file, despite the tool is restricted and has it's visibility set to private
Impact
This issue may lead to restricted tools being invoked by users via the chat completion API
Resolution
Fixed across two releases
-
Local Tool records (
tool ids: ["<tool id>"]referencing a stored Tool): fixed in commit 9b06fdc8f, first released in v0.7.0.get tools()inbackend/open webui/utils/tools.py(line 166) now resolves the caller's group memberships and rejects each requestedtool idwhose owner isn't the caller and for which noAccessGrants.has access(resource type='tool', permission='read')grant exists. Admins continue to bypass whenBYPASS ADMIN ACCESS CONTROLis enabled (its documented UI/posture purpose). -
Admin-configured MCP servers (
tool ids: ["server:mcp:<id>"], the report's exact PoC): fixed in commit 4737e1f11, first released in v0.8.6. The MCP-resolution loop inbackend/open webui/utils/middleware.py(line 2670) now callshas connection access(user, mcp server connection)and skips the server with a warning if the user has no grant — the server's stored credentials are never used on behalf of an unauthorized caller.
Users on
>= 0.8.6 are not affected.Fix
Found an issue in the description? Have something to add? Feel free to write us 👾
Related Identifiers
Affected Products
Open-Webui