Global Knowledge Base Enumeration via knowledge-bases Meta-Collection

Affected Component

Retrieval collection access validation:

• backend/open_webui/routers/retrieval.py (lines 2330-2355, _validate_collection_access)
• backend/open_webui/routers/retrieval.py (query endpoints, e.g. POST /query/doc)

Affected Versions

Current main branch (commit 6fdd19bf1) and likely all versions with the knowledge base subsystem.

Description

The _validate_collection_access function uses an incomplete allowlist that only enforces ownership checks for collections matching user-memory-* and file-* patterns. All other collection names pass through unchecked — including the system-level knowledge-bases meta-collection, which stores the IDs, names, and descriptions of every knowledge base on the instance.

Any authenticated user can query this meta-collection directly via the retrieval query endpoints to obtain a global index of all knowledge bases across all users.

# retrieval.py:2330-2355 — incomplete collection allowlist
def _validate_collection_access(user, collection_name, ...):
    if collection_name.startswith('user-memory-'):
        # Check user-memory ownership
        ...
    elif collection_name.startswith('file-'):
        # Check file access
        ...
    # Everything else (including "knowledge-bases") passes through unchecked

This finding is the enabler for the KB destruction (process/web), KB content injection (process/file), and RAG vector search access bypass findings — all of which require knowing a target KB's UUID. Without this enumeration, UUIDs are random and practically unguessable; with it, UUIDs across the entire instance are trivially obtained.

CVSS 3.1 Breakdown

Attack Scenario

1. Attacker (any authenticated user) sends:

   POST /api/v1/retrieval/query/doc
   {
     "collection_name": "knowledge-bases",
     "query": "confidential"
   }
   

2. _validate_collection_access does not recognize the knowledge-bases prefix and lets the request pass.
3. The vector search returns the most relevant documents from the meta-collection — knowledge base records including their UUIDs, names, and descriptions — across all users on the instance.
4. Attacker varies the query to enumerate more KBs: "project", "internal", "private", etc.
5. Attacker now has a full target list for subsequent attacks (destruction, poisoning, content extraction).

Impact

• Information disclosure: KB names and descriptions may reveal sensitive project names, internal initiatives, or user activities
• Enabler for other attacks: Unlocks the following findings by supplying the required target UUIDs:
  • KB destruction/poisoning via process/web
  • Cross-user content injection via process/file
  • RAG vector search access bypass in retrieval/utils.py
• Transforms these from theoretical (requires UUID guessing) to trivially exploitable (UUIDs enumerable)

Preconditions

• Attacker must have a valid user account