CVE-2026-58076

CVE-2026-58076 is a deserialization flaw in Apache Airflow's executor_config handling, but framing it as a simple code bug misses the structural problem. Airflow's security model explicitly states that Dag authors are untrusted — yet executor_config, a field Dag authors populate, flows through the serialization layer into import_string() calls that execute arbitrary code. The patch for CVE-2026-33264 addressed the happy path through the deserializer but left the exception branch intact. That gap is now CVE-2026-58076. The BaseException allowlist fix is reasonable but raises a deeper question: why was dynamic import needed here at all? Exception reconstruction could use a registry or allowlist. The architectural choice to use import_string() with no constraints created the conditions for multiple CVEs in the same code path. The EPSS score of 0.0028 will mislead triage teams. The trigger requires no authenticated request and no user interaction — the Scheduler deserializes Dag files as part of its normal loop. Any Dag author with write access to the Dag repository (a standard workflow, not elevated privilege) can trigger this by pushing a malicious Dag file. The Scheduler then executes the payload on its own cadence with metadata database credentials and JWT secrets already in memory. Check your Airflow deployment: if shared Dag storage or git-sync is in use, the attack surface includes anyone who can push to that repository. Prioritize patching, restrict Dag file push access to the minimum required, and audit executor_config usage in your Dags to identify any custom classes that might conflict with the BaseException restriction. The long-term question is whether Airflow's security model can realistically hold given how executor_config works — that architectural review may be unavoidable.

Reviewed through automated stages and approved by a human before publication.