CloudNativePG EUVD-2026-32930

Impact

The CloudNativePG metrics exporter opens its PostgreSQL connection as the postgres superuser via the pod-local Unix socket, then demotes the session with SET ROLE pg_monitor. SET ROLE changes only current_user; session_user remains postgres. That residual superuser identity is the foothold for the rest of the chain.

Any SQL expression evaluated inside the scrape session can invoke RESET ROLE to recover real superuser privileges, then use COPY ... TO PROGRAM to spawn an OS-level subprocess as the postgres user inside the primary pod. The READ ONLY transaction flag does not block this; it gates writes to database state, not external processes.

Two exploitation paths follow from this root cause.

Path 1: custom metric queries with unqualified identifiers (all supported releases)

A database user who owns a schema on the search_path of any scraped database can plant a shadow object whose name matches an unqualified identifier in a custom metric query. When the exporter next evaluates that query, the shadow expression executes inside the session_user = postgres scrape session, giving the attacker PostgreSQL superuser privileges and OS command execution inside the primary pod within one scrape interval (≤30 s). Exploitability requires a custom metric query that contains an unqualified relation or function reference.

Although search_path shadowing of unqualified identifiers is the most direct case, the underlying bug is that any expression evaluated inside the scrape session is a superuser code path. Other exploitable shapes include user-defined functions, operators or casts resolved during the scrape, joins or subqueries against user-owned tables and views, and index expressions or RLS policies on read-touched objects.

Path 2: stock default-monitoring.yaml (all supported releases, no custom metrics required)

The pg_extensions metric shipped in default-monitoring.yaml used an unqualified current_database() call and ran against every user database (target_databases: '*'). Any non-superuser who owns a user database (including the default app role created by bootstrap.initdb) could shadow current_database() and trigger the full escalation chain against a stock CNPG deployment on the first scrape after the shadow was planted.

Combined impact

The chain yields privilege escalation from a low-privileged database role (e.g. the default app role) to PostgreSQL superuser, plus arbitrary OS command execution as the postgres user inside the primary pod, all within one scrape interval. A web application SQL injection vulnerability in an app backed by a CNPG cluster is therefore sufficient to pivot to database-pod RCE.

Who is impacted

• All deployments on any supported release with default monitoring enabled are affected by Path 2.
• All deployments on any supported release that use custom metric queries containing unqualified catalog references are affected by Path 1.
• Multi-tenant platforms that allow customers to supply or influence custom metric query bodies are at the highest risk for Path 1.

Patches

Three separate patches address the vulnerability.

Patch 1: PR #10576 "schema-qualify catalog references in default monitoring queries and documentation samples"

Schema-qualifies all unqualified pg_catalog function and view references in the shipped default-monitoring.yaml and in documentation examples. This closes Path 2 in operator-shipped configuration and removes the unqualified-identifier attack surface from all operator-shipped metric queries. Operators who clone or copy default-monitoring.yaml into custom monitoring ConfigMaps, or have copy-pasted unqualified queries elsewhere, must re-qualify those queries themselves.

Backported to all currently supported releases:

• v1.29.x (x ≥ 1)
• v1.28.x (x ≥ 3)

Patch 2: "dedicated cnpg_metrics_exporter role with pg_ident.conf peer mapping"

Introduces a dedicated cnpg_metrics_exporter PostgreSQL role (granted pg_monitor, no superuser privileges) and maps it in pg_ident.conf via peer authentication on the local Unix socket, following the same pattern already used for cnpg_pooler_pgbouncer. The metrics exporter connects as this role instead of postgres, so session_user is never a superuser and RESET ROLE has no escalation effect. This eliminates the root cause entirely.

Demoting the session at the SQL level (via SET SESSION AUTHORIZATION pg_monitor) is not sufficient: the privilege check for SET SESSION AUTHORIZATION is whether the *authenticated* user is a superuser, not the current session_user. With the connection still authenticated as postgres, any SQL in the session can run RESET SESSION AUTHORIZATION and recover the original superuser identity. This is the same recovery primitive as RESET ROLE, one layer up. Only changing the authenticated user closes the loop.

With this change in place, the original chain breaks at every step: RESET ROLE and RESET SESSION AUTHORIZATION cannot recover superuser, and COPY ... TO PROGRAM requires a privilege pg_monitor does not grant. As defense in depth, the monitoring transaction also prepends pg_catalog to the connection's search_path, so unqualified catalog identifiers cannot resolve to user-planted shadow objects.

This patch changes the connection identity but not how queries are evaluated. Custom metric queries within pg_monitor's scope (catalog reads, pg_stat_* views, settings) continue to work without modification. Queries that previously relied on superuser-level access (reading user-owned tables not granted to cnpg_metrics_exporter, or superuser-only catalogs such as pg_authid or pg_subscription) will fail and need explicit GRANT statements to cnpg_metrics_exporter.

The role is created and maintained with PASSWORD NULL; any password set out-of-band is cleared on the next reconcile, so the role cannot be authenticated by password regardless of operator pre-creation.

For replica clusters, upgrade the source primary cluster before any replica clusters that consume from it. The cnpg_metrics_exporter role is created on the source primary and replicates downstream; a replica cluster upgraded first will scrape against a missing role until the source primary upgrades or the role is created manually (see the monitoring documentation).

The patch will be backported to all currently supported releases:

• v1.29.x (x ≥ 1)
• v1.28.x (x ≥ 3)

Workarounds

If upgrading immediately is not possible:

1. Schema-qualify all identifiers in custom metric queries. Use explicit pg_catalog. prefixes for all catalog functions and views (e.g. pg_catalog.current_database(), pg_catalog.now()). This is a partial mitigation: it closes the search_path-shadowing shape in operator- and user-supplied metric bodies, but other expression shapes (user-defined functions, operators or casts; joins or subqueries on user-owned tables and views; RLS policies on read-touched objects) remain superuser code paths until Patch 2 lands.
2. Restrict database ownership. Ensure only fully trusted roles own user databases in scraped clusters. The exploit requires the ability to plant an object on the metrics exporter's search_path in a scraped database, typically by owning the database (and therefore public via pg_database_owner) or by holding CREATE on a schema already reachable through search_path.

*PG <15 caveat:* public grants CREATE to PUBLIC by default before PostgreSQL 15, so any authenticated role in a scraped database can plant a shadow object regardless of ownership.

1. Limit the scope of target_databases: '*' queries. Avoid target_databases: '*' unless every database in the cluster, and every role that owns one, is fully trusted. Where possible, restrict target_databases to specific, known-safe databases.
2. Do not expose metric query SQL to untrusted users. Multi-tenant platforms that allow customers to supply or influence custom metric query bodies should treat this as a critical trust boundary until the architectural fix is released.

References

• Fix (Patch 1): PR #10576 "schema-qualify catalog references in default monitoring queries and documentation samples"
• Fix (Patch 2): "dedicated cnpg_metrics_exporter role with pg_ident.conf peer mapping"
• Reported by: Mehmet Ince