Snowflake Spark Connector CVE-2026-15183

Exploitation requires that the vulnerable spark-snowflake connector (pre-3.2.1) be in use and, for the highest-impact SQL injection and privilege-escalation paths, that the attacker have access to a SHARED Spark or Spark-SQL session where they can issue runtime SET commands or supply staging/connector options that feed the SnowflakeFallbackCatalog option map - which then executes under the cluster admin's JDBC credentials. The OAuth credential-theft path requires the ability to influence the OAuth token-request URL, and the COPY-redirection path requires the ability to place malicious files in an ingestion pipeline or control staging options. Limiting factors: single-tenant clusters with no untrusted session participants remove the shared-session injection vectors; the CVSS AC:H and AT:P metrics indicate meaningful attack complexity and a required precondition; and the value of injected SQL depends on the (ideally least-privilege) Snowflake role the JDBC session runs as.

The CVSS 4.0 base score of 9.2 (Critical) reflects network vector with full high impact to confidentiality, integrity, and availability of the vulnerable system (VC:H/VI:H/VA:H), but two metrics temper real-world ease: AC:H (high attack complexity) and AT:P (a present attack requirement/precondition). Although PR:N is asserted, the description's most concrete attack paths - injecting via staging options 'in a shared Spark environment' and issuing runtime SET commands 'in a shared Spark-SQL session' - imply the attacker already has access to submit jobs or queries into a shared cluster, so effective privileges are higher than PR:N suggests. No EPSS score, KEV listing, POC, or SSVC decision data was provided, so exploitation probability cannot be quantified. Overall this is a genuine high-priority issue for organizations running multi-tenant Spark clusters with this connector, but likely a lower practical risk for single-tenant deployments where no untrusted party shares the Spark session - verify your deployment model. Note: CVSS asserts network attack vector while the described injection paths imply access to a shared Spark session; reconcile with vendor guidance.

In a shared Spark-SQL cluster, a low-privileged analyst issues a runtime SET command whose value contains crafted SQL; the connector folds this into the SnowflakeFallbackCatalog option map and executes it over the cluster admin's JDBC connection, running attacker SQL with the admin's Snowflake role. Alternatively, the attacker supplies malicious staging options or a crafted OAuth token-request URL to steal OAuth client credentials or redirect a COPY operation to storage they control. No public exploit is identified at time of analysis, and the high attack complexity plus required access to the shared session limit who can carry this out.