GHSA-34V6-3CHG-HCJQ
Two authentication-state hardening items: MCP tool refusals let an authenticated caller probe which databases exist, and a failed Bolt LOGON leaves the previous principal bound
Ciclo de vida de la vulnerabilidad
| Fecha | Fase | Evento | Detalle |
|---|---|---|---|
| Divulgación | Publicada | GHSA-34V6-3CHG-HCJQ divulgada | |
| Monitoreo | Detectada | Ingresó al índice | |
| Monitoreo | Verificada | Última verificación contra las fuentes |
Two low-severity items from a routine audit of main at `888b83c926`. Both reproduced. Reported privately per SECURITY.md even though neither gives an attacker access they did not already have, because both break an invariant the codebase states elsewhere and keeps.
---
1. MCP tool refusals distinguish "forbidden" from "does not exist"
An MCP caller allowed onto the surface at all can enumerate the names of databases it is not granted, one guess at a time, because the two refusals differ. The "does not exist" arm also prints the list of databases the caller can reach, which is fine on its own.
The rule is stated twice in the codebase and kept in both places. `MCPResources.read` (mcp/src/main/java/com/arcadedb/mcp/MCPResources.java:67-78) says: "An unknown database, an unauthorized database, and a malformed URI are indistinguishable to the caller, so that this method cannot be used to probe which databases exist". The Bolt listener states the same rule for the same reason (bolt/src/main/java/com/arcadedb/bolt/BoltNetworkExecutor.java:1394-1398).
The tool path does not keep it. `MCPToolUtils.resolveDatabase` (mcp/src/main/java/com/arcadedb/mcp/tools/MCPToolUtils.java:60-84) tests `existsDatabase` at :70 before `canAccessToDatabase` at :79, and the two arms raise different messages. `resolveDatabase` is the entry point of `query`, `execute_command`, `get_schema`, `sample_records`, `vector_search`, `full_text_search`, `hybrid_search`, `upsert_entity` and `upsert_relationship`, so the oracle is on nine of the sixteen tools.
Both messages reach the client verbatim in every server mode: `MCPDispatcher`'s `SecurityException` and `MCPToolArgumentException` arms both answer through `ErrorConcealment.describe`, which returns the message unchanged (server/src/main/java/com/arcadedb/server/ErrorConcealment.java:110-115).
Reproduction
A user granted only the test database, calling with `allowedUsers ["*"]`, the documented agent-delegation configuration:
PROBE-MCP query('auditr25secret') -> SecurityException: User 'auditr25limited' is not authorized to access database 'auditr25secret'
PROBE-MCP query('auditr25doesnotexist') -> MCPToolArgumentException: Database 'auditr25doesnotexist' does not exist. Available databases: [graph]. Use one of these names or call list_databases to refresh the list.
PROBE-MCP resources/read('auditr25secret') -> MCPResourceNotFoundException: Resource not found: arcadedb://auditr25secret/schema
PROBE-MCP resources/read(missing) -> MCPResourceNotFoundException: Resource not found: arcadedb://auditr25doesnotexist/schemaThe last two lines are the contrast: `resources/read` answers identically either way, which is what the tool path should do.
Refutation attempted
Is the "does not exist" message a deliberate LLM self-correction aid? The javadoc at `MCPToolUtils.java:55-59` says so for the list of available databases, which is fine because it lists only what the caller may reach. It says nothing about leaking the existence of a database the caller may not reach, and the two neighbouring surfaces state the opposite rule.
Fix, and the guarantee
Move the authorization test above the existence test, as Bolt does, and answer one message for both.
The guarantee: a shared helper that resolves a database name for every MCP surface including `resources/read`, so the ordering is stated once; plus a test asserting the refusal for an existing-but-forbidden name is byte-identical to the refusal for a name that does not exist, enumerating the tools from `MCPDispatcher`'s own registry rather than from a list.
---
2. A failed Bolt LOGON leaves the previous principal bound
`handleLogon` accepts a LOGON from READY as well as from AUTHENTICATION (bolt/src/main/java/com/arcadedb/bolt/BoltNetworkExecutor.java:589-593), which is what makes LOGOFF-less re-authentication possible on a pooled connection. `authenticateUser` (:1830-1852) sets FAILED on a refusal but does not clear `user` and does not call `markUnauthenticated` (:1879). `handleReset` (:738-760) then restores READY whenever `user` is non-null.
Net effect: an attempt to switch the connection's identity that fails leaves the old identity in force and serving.
The invariant is stated, and kept, by the twin on the Redis wire. `RedisNetworkExecutor.markUnauthenticated` (redisw/src/main/java/com/arcadedb/redis/RedisNetworkExecutor.java:1048-1060) has a javadoc saying a connection that "fails a subsequent AUTH/HELLO re-authentication attempt on the same connection" must have `authenticatedUser` go back to null. The Bolt executor has the same method, written against the same issue, and calls it only from `handleLogoff` (:636).
Reproduction
Driving the real handlers:
PROBE-LOGON before: state=READY user=root
PROBE-LOGON after failed LOGON: state=FAILED user=root
PROBE-LOGON after RESET: state=READY user=rootRefutation attempted
Exploitable by an unauthenticated attacker? No. The connection must already be authenticated as the privileged principal, so nobody gains access they did not have. That is why we rate it low.
Does the official Neo4j driver reach it? Its re-authentication sends LOGOFF then LOGON, and after LOGOFF `user` is null, so the driver's own path is safe. The exposure is a proxy or pool that sends LOGON from READY without LOGOFF, which the server accepts.
Does FAILED contain it? Only until the client RESETs, which is exactly what a pool does before reusing a connection, and RESET is what restores READY with the stale principal.
Fix, and the guarantee
Clear `user` and call `markUnauthenticated` on every refusal inside `authenticateUser`, so "authentication did not succeed" always implies "no principal is bound and the bounded pre-auth read timeout is armed".
The guarantee: make the binding unrepresentable without a successful authentication. Have `authenticateUser` null `user` first and assign it only from a successful result, rather than leaving the previous value in place when the assignment does not happen. A test should assert over the pair (user, socket timeout) after a refusal, for a connection that was authenticated and one that was not, since asserting only on the state enum passes on the current code.
---
Dedup
Nothing matching either item in the last 1000 issue titles. #9772 is about MCP argument refusals being concealed, the opposite direction. #6803 is Bolt LOGOFF leaving a transaction and a result set pinned, which is fixed; this is the adjacent field on the adjacent handler on the failure path.
Incluye detalles técnicos publicados por el investigador. Úselos solo para validar y corregir sus propios sistemas.
Rangos afectados
Estado del parche por paquete
| Paquete / producto | Estado | Corregida en | Desde divulgación | Última versión |
|---|---|---|---|---|
Mavencom.arcadedb:arcadedb-mcp |
Sin parche | — | — | — |
Arsenal público
Sin exploits, módulos ni PoC públicos conocidos a la fecha de verificación.
https://github.com/ArcadeData/arcadedb/security/advisories/GHSA-34v6-3chg-hcjqLas referencias externas se muestran para copiar: cve.llama.ec no lo redirige a otros sitios.
