Executive brief
Apache Airflow is a workflow orchestration platform used to schedule and monitor data pipelines. A vulnerability in its Task SDK allows a DAG author to cause arbitrary Python modules to be imported within the scheduler process by crafting malicious serialized callback objects. An attacker exploiting this could execute arbitrary code on the scheduler, potentially compromising the entire orchestration infrastructure and any downstream systems it controls.
Technical details
The vulnerability is a deserialization attack in Apache Airflow's Task SDK. When deserializing a `Callback` object from stored `next_kwargs`, the SDK reconstructs the callback by re-running its constructor, which imports the module specified in the stored callback path. The `SyncCallback` class passes the default `allowed_deserialization_classes` allow-list, so standard mitigation measures do not prevent this attack. A DAG author (who controls a task instance's `next_kwargs` through the task execution API) can inject a malicious callback path that triggers arbitrary module imports during the scheduler's `awaiting_input` timeout sweep. The sweep runs unconditionally and requires no non-default configuration. This is a distinct gadget from CVE-2026-58076 and CVE-2026-67260, affecting only Airflow 3.0.0 through 3.3.0.
Affected products
- Apache Airflow 3.0.0 through 3.3.0
Timeline
- 2026-08-12: disclosed
- 2026-08-04: patched: Fix merged in PR #70704; upgrade to 3.3.1 or later