Executive brief
Apache Airflow is a workflow orchestration platform that manages scheduled data pipeline jobs. A critical flaw in how Airflow deserializes DAG (Directed Acyclic Graph) configuration files allows an attacker who can author DAGs to inject malicious code that executes on the API Server or Scheduler—core processes that should only run trusted Airflow platform code. This completely breaks Airflow's security model and gives attackers full control over the infrastructure running the platform.
Technical details
The vulnerability is a deserialization flaw (CWE-502) in Apache Airflow's BaseSerialization.deserialize() method. When loading serialized DAGs, the code performs unrestricted dynamic imports via import_string() on attacker-controlled class paths embedded in trigger_kwargs. A DAG author can craft a malicious serialized trigger object containing a class path pointing to arbitrary code; when the Scheduler or API Server deserializes this DAG, the attacker's code executes in the context of those privileged processes. This violates Airflow's core security boundary—DAG author code should never run in the Scheduler or API Server. The vulnerability requires no authentication or user interaction; it is triggered automatically during normal DAG loading. Patches were released in Apache Airflow 3.3.0 and backported via commits 69b6c54 and 7e9bdb2, which skip deserialization of trigger_kwargs. Defense-in-depth mitigation via the [core] allowed_deserialization_classes config can further restrict dynamic imports to a whitelist.
Affected products
- Apache Airflow < 3.3.0
Timeline
- 2026-07-07: disclosed: Vulnerability published by Apache and NVD
- 2026-05-07: patched: Fix merged in PR #66002 (commit 69b6c54)
- 2026-06-17: patched: Cleanup merged in PR #68528 (commit 7e9bdb2)