[Security/Medium] readPrivate dateOfBirth-search gate bypassable via Lucene query_string #1724

Severity: Medium — Confidence: 0.85

Location: backend/social-work-app/lambdas/python/search/handlers/search.py:204-258 (_query_references_field / _validate_date_of_birth_permission). Identical vulnerable copy at backend/cosmetology-app/lambdas/python/search/handlers/search.py:204-258.

Not affected: compact-connect (production) — its search handler has no such gate and does not index dateOfBirth into OpenSearch, so this vector does not apply there.

Summary

The advanced search endpoint accepts a raw OpenSearch DSL query (SearchProvidersRequestSchema.query = Raw, only index/_index keys blocked) and is authorized only by @authorize_compact_level_only_action(READ_GENERAL). Searching on dateOfBirth (indexed as date, opensearch_client.py:282) is meant to require the higher readPrivate scope. The sole enforcement, _query_references_field, only flags a dict key ending in .dateOfBirth or a string value that equals/ends-with .dateOfBirth — it never inspects substrings inside a string value.

Exploit chain

A staff user with only {compact}/readGeneral (implicitly held by all staff), not readPrivate, POSTs to /v1/compacts/{compact}/providers/search:

{"query":{"bool":{"must":[
  {"term":{"providerId":""}},
  {"query_string":{"query":"licenses.dateOfBirth:[1985-01-01 TO 1985-06-30]"}}
]}}}

The value "licenses.dateOfBirth:[...]" ends with ], not .dateOfBirth, so the gate passes and the query runs. The response schema doesn't echo DOB, but hit-vs-no-hit + total is a boolean oracle — binary-searching the date range (~15 requests) yields any provider's exact DOB, defeating the readGeneral/readPrivate boundary. wildcard/regexp/script types are also unrestricted.

Reachability

Wired & real on main. Route added in stacks/search_api_stack/v1_api/provider_search.py (add_method('POST',...)), scope ['{compact}/readGeneral'] (v1_api/api.py:29-40); SearchApiStack instantiated at pipeline/backend_stage.py:197 for both apps; dateOfBirth is in the index mapping (type: date).

Suggested fix

Don't rely on field-name string matching over attacker-controlled DSL.

• Preferred: allowlist safe query types (term/terms/match/bool/nested/range) and reject free-form query_string/simple_query_string/wildcard/regexp/script.
• Minimal stopgap: match the dateOfBirth substring inside any string value and reject fields specs containing *.
• Best long-term: OpenSearch field-level masking.

Apply to both social-work-app and cosmetology-app copies.