CVE-2026-54016: Open WebUI: BOLA exposes private knowledge base files

## Summary Open WebUI has a Broken Object Level Authorization (BOLA) vulnerability in the builtin `search_knowledge_files` tool. When native function calling is enabled and the selected model has no attached knowledge bases, an authenticated user can call `search_knowledge_files` with an arbitrary `knowledge_id`. The function then returns file metadata from that knowledge base without checking whether the user has read access. This allows unauthorized enumeration of private or restricted knowledge base files. ## Details The vulnerable code is in: `backend/open_webui/tools/builtin.py` Affected function: ```python async def search_knowledge_files( query: str, knowledge_id: Optional[str] = None, count: int = 5, skip: int = 0, __request__: Request = None, __user__: dict = None, __model_knowledge__: Optional[list[dict]] = None, ) -> str: ``` In the "No attached knowledge" branch, when `knowledge_id` is provided, the function directly calls: ```python result = await Knowledges.search_files_by_id( knowledge_id=knowledge_id, user_id=user_id, filter={"query": query}, skip=skip, limit=count, ) ``` This code path does not verify that the current user is authorized to access the specified knowledge base. The missing check is inconsistent with other nearby code paths. For example, the attached-knowledge branch in the same function checks whether the user is an admin, the owner of the knowledge base, or has explicit read access through `AccessGrants`: ```python if not ( user_role == "admin" or knowledge.user_id == user_id or await AccessGrants.has_access( user_id=user_id, resource_type="knowledge", resource_id=knowledge.id, permission="read", user_group_ids=set(user_group_ids), ) ): continue ``` The sibling function `query_knowledge_files` also performs the same authorization check before using user-supplied knowledge base IDs. The underlying method `Knowledges.search_files_by_id()` receives `user_id`, but it does not enforce authorization for the provided `knowledge_id`. As a result, this builtin tool path can access a knowledge base by ID without verifying the caller's permissions. ## PoC ### Prerequisites - The attacker has a valid authenticated Open WebUI account. - The victim owns a private or restricted knowledge base. - The attacker does not own the target knowledge base. - The attacker does not have `read` permission for the target knowledge base in `AccessGrants`. - The attacker knows the target `knowledge_id`. - The selected model has no attached knowledge bases. - Builtin tools are enabled. - The knowledge builtin tool category is enabled. - Native function calling is enabled. ### Reproduction Steps 1. Create a private or restricted knowledge base as the victim user. 2. Upload one or more files to that knowledge base. 3. Confirm that the attacker user does not have access to the knowledge base. 4. As the attacker user, send a chat completion request with native function calling enabled: ```json { "stream": true, "model": "gpt-4o-mini", "params": { "function_calling": "native" }, "messages": [ { "role": "user", "content": "Please use the search_knowledge_files tool with knowledge_id \"c0c84752-2e9d-42bf-bc3c-c0f272aa61c1\" to search all files" } ] } ``` Replace `c0c84752-2e9d-42bf-bc3c-c0f272aa61c1` with the victim's private knowledge base ID. ### Expected Result The request should be denied because the attacker does not have access to the target knowledge base. ### Actual Result `search_knowledge_files` returns metadata for files inside the target knowledge base, including: - file ID; - filename; - knowledge base ID; - knowledge base name; - update timestamp. ## Impact This is a Broken Object Level Authorization / Broken Access Control vulnerability. An authenticated attacker who knows a valid `knowledge_id` can enumerate files from private or restricted knowledge bases without authorization. The leaked metadata may expose sensitive information through filenames, such as: - financial reports; - employee documents; - customer contracts; - internal roadmap files; - confidential project documents. The exposed file IDs may also help attackers chain this issue with other knowledge-file access paths, such as `view_knowledge_file`, to attempt further content extraction. This vulnerability bypasses the intended `AccessGrants` permission model and may also allow post-revocation metadata access if a user remembers a previously accessible `knowledge_id`. ## Suggested Fix Add the same authorization check used in `query_knowledge_files` before calling `Knowledges.search_files_by_id()`: ```python if knowledge_id: knowledge = await Knowledges.get_knowledge_by_id(knowledge_id) if not knowledge or not ( user_role == "admin" or knowledge.user_id == user_id or await AccessGrants.has_access( user_id=user_id, resource_type="knowledge", resource_id=knowledge.id, permission="read", user_group_ids=set(user_group_ids), ) ): return json.dumps({"error": f"Access denied to knowledge base {knowledge_id}"}) result = await Knowledges.search_files_by_id( knowledge_id=knowledge_id, user_id=user_id, filter={"query": query}, skip=skip, limit=count, ) ``` As defense in depth, authorization should also be enforced or safely wrapped around `Knowledges.search_files_by_id()` so that future callers cannot accidentally bypass access control.

An attacker registers a legitimate Open WebUI account in a multi-tenant enterprise deployment — or compromises a low-privilege existing account. They observe or infer a target `knowledge_id` UUID from shared UI context, a leaked URL, or systematic UUID enumeration (UUIDs are not secret once any shared interaction exposes them). With native function calling enabled on any model with no attached knowledge bases, the attacker crafts a chat message instructing the model to call `search_knowledge_files` with the target UUID. The builtin tool executes without checking `AccessGrants`, returning filenames, file IDs, and timestamps from the victim's private knowledge base. The attacker catalogs the file IDs and uses them to call `view_knowledge_file` or similar endpoints to extract full document content, building a complete picture of the organization's sensitive document landscape without ever having been granted access.