Skip to main content
Menu
Advisories

arcadedb-native protection bypass

GHSA-WW64-52X2-W3X2Published Updated GitHub

Description

Addendum: scope of the native-image read-only bypass, measured with a differential (native 26.10.2 vs JVM 26.10.2, same statements, same HTTP door)

The advisory covers SELECT (INSERT ...) FROM T and SELECT ... LET $y = (INSERT ...). A systematic run of 210 statement shapes through POST /api/v1/query/{db} (SQL, openCypher via both opencypher and cypher, sqlscript, the GET form, three serializers), each on a database reset to the same state first, shows the hole is the whole nesting surface and not two shapes. The JVM refuses every one with QueryNotIdempotentException (400).

  1. Accepted by the read-only door on native and refused on JVM: 46 of the 177 shapes in ro2, 18 of 33 in ro3. Durable state changed after the call (row count, sum, type list and index count read back through POST /command): 38 in ro2, 3 in ro3. Every insert-class form persisted: INSERT or CREATE VERTEX inside a SELECT projection, a LET right-hand side (also the second LET of a pair, a nested SELECT FROM (SELECT FROM (INSERT ...))), a FROM subquery, expand((INSERT ...)), a WHERE or ORDER BY subquery, a function argument (coalesce((INSERT ...), 1)), a MATCH where: and RETURN subquery, a TRAVERSE start and WHILE, a sqlscript LET $x = (INSERT ...), the GET form GET /api/v1/query/{db}/sql/{statement}, and the record, graph and studio serializers.
  2. Top-level PROFILE INSERT INTO T SET a = 106 and PROFILE CREATE VERTEX T SET a = 108 run and persist on native (JVM: 400). This is the #9313 shape (PROFILE classified as READ whatever it wraps), which the advisory does not mention: the PROFILE and EXPLAIN wrappers hold the inner statement in a field the reflective walk cannot read. EXPLAIN INSERT does not execute on either server.
  3. DML and DDL forms (UPDATE, DELETE, TRUNCATE, CREATE/DROP TYPE, ALTER, CREATE INDEX, CREATE EDGE, REBUILD INDEX, CHECK DATABASE, ALIGN DATABASE, MOVE VERTEX, BEGIN) inside the same positions are accepted by the door as well; in 10 + 34 cases they then fail with 500 UnsupportedOperationException: Implement UpdateStatement.refersToParent() (and the matching name for each statement class) instead of the 400, and no state change was observed. In FROM (UPDATE|DELETE|...) and PROFILE UPDATE|DELETE|DROP TYPE with matching rows the call returned 200 and the state read back unchanged. We did not establish why only insert-class forms persisted, so do not rely on the other forms being harmless: the classification is wrong for all of them.
  4. The same counts (46 accepted, 38 durable in ro2; 18 accepted, 3 durable in ro3) appear on arcadedata/arcadedb:26.11.1-SNAPSHOT-native-amd64 (main, built 2026-10-21 from 7050ef7), so the hole is still open on main.
  5. A side effect that makes it more visible: SELECT FROM T WHERE $parent.$current.a IN (SELECT a FROM (INSERT INTO T SET a = 34)) inserts one row per outer row; repeated nested inserts in the earlier cascading run grew T from 2 rows to more than 48,000 (RIDs up to #1:48639).

Exposure is every gate that consumes the classification (read-only HTTP query door, the GET query route, MCP read-only mode, HA leader routing, requireStreamableStatement), as in the advisory text. Fix guidance from the advisory still applies (generated allDeclaredFields metadata or a reflection-free walk, plus a native smoke test that POSTs one nested write to /api/v1/query); the test list should include PROFILE INSERT, a FROM subquery and a WHERE subquery, not only the two LET/projection shapes.

Statements whose effect persisted on native (ro2: 38 of 46 accepted; JVM refused every one with 400)

  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T
  • SELECT count(*) AS n, sum(a) AS s FROM T

ro3: shapes where the state after the call differs between native and JVM (4: the three accepted with 200 that persisted, plus one nested CREATE EDGE that answered 500 TransactionException on native and still left its edge in place)

statement sent to POST /api/v1/querynative statusJVM statusstate after on nativestate after on JVM
PROFILE CREATE VERTEX T SET a = 108200400 QueryNotIdempotentException3 rows/sum 111, types 4, idx 0, edges 12 rows/sum 3, types 4, idx 0, edges 1
SELECT FROM (CREATE VERTEX T SET a = 110)200400 QueryNotIdempotentException3 rows/sum 113, types 4, idx 0, edges 12 rows/sum 3, types 4, idx 0, edges 1
SELECT FROM (CREATE EDGE E FROM (SELECT FROM T WHERE a = 1) TO (SELECT FROM T WHERE a = 2))500 TransactionException400 QueryNotIdempotentException2 rows/sum 3, types 4, idx 0, edges 22 rows/sum 3, types 4, idx 0, edges 1
SELECT FROM (INSERT INTO T SET a = 111 RETURN a)200400 QueryNotIdempotentException3 rows/sum 114, types 4, idx 0, edges 12 rows/sum 3, types 4, idx 0, edges 1

Reproduce

docker run -d --name ro -p 2480:2480 -e JAVA_OPTS= -e arcadedb.server.rootPassword=playwithdata arcadedata/arcadedb:26.10.2-native-amd64
curl -s -u root:playwithdata -X POST localhost:2480/api/v1/server -d '{"command":"create database insp"}'
curl -s -u root:playwithdata -X POST localhost:2480/api/v1/command/insp -d '{"language":"sql","command":"CREATE DOCUMENT TYPE T"}'
curl -s -u root:playwithdata -X POST localhost:2480/api/v1/query/insp -d '{"language":"sql","command":"PROFILE INSERT INTO T SET a = 106"}'   # native 200, JVM 400 QueryNotIdempotentException
curl -s -u root:playwithdata -X POST localhost:2480/api/v1/query/insp -d '{"language":"sql","command":"SELECT count(*) AS c FROM T"}'   # native 1

Same counts on arcadedata/arcadedb:26.11.1-SNAPSHOT-native-amd64 (built 2026-10-21). Found with a native-vs-JVM differential of 51,086 statements over HTTP; everything else agreed apart from known or documented gaps.

Related advisories

This issue is tracked together with GHSA-84m7-76gh-rm42, which carries the CVE identifier for the group.

Affected versions

References

GitHub · ArcadeData/arcadedb