GPU VulnDB

Database/Control plane, storage & DevOps

Apache Airflow: DAG authors can execute code in the scheduler through the doc_md parameter

CVSS 8.8CVE-2024-39877Control plane, storage & DevOpscurated

Impact

A DAG author - a role the Airflow security model treats as less privileged than a deployment admin - can place a crafted doc_md value in a DAG file and have it evaluated as code inside the scheduler process. On a shared training or data-prep cluster, the scheduler usually runs with the broadest credentials in the deployment: connection secrets, the metadata database, cloud and registry credentials used to launch GPU jobs. Code execution there is a pivot from "can submit a pipeline" to "owns the orchestration layer", including the ability to redirect other tenants' jobs. Affects 2.4.0 through versions before 2.9.3.

Who can reach it

Authenticated user with DAG-author privileges (ability to add or modify a DAG file the scheduler parses). No admin rights needed; not reachable by an unauthenticated network caller.

What to do

Upgrade apache-airflow to 2.9.3 or later and restart the scheduler, webserver and workers. No node reboot or firmware work; the cost is an orchestration-layer restart window, during which queued GPU jobs are not scheduled. Until then, treat DAG-file write access as equivalent to scheduler code execution and restrict who holds it.

References

Related entries

All Control plane, storage & DevOps entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.