CVE-2026-129509 affects Apache Airflow Microsoft Azure provider versions 10.2.0 through 15.3.0, where S3ToAzureBlobStorageOperator set s3_verify=False by default, disabling TLS certificate verification for the S3 connection regardless of AWS connection settings. A network-positioned attacker could intercept and tamper with that traffic.
Apache Airflow before 3.4.0 revokes the presented token on logout only best-effort: failures are silently ignored and logout still reports success. When auth managers redirect logout to an external URL such as FAB or Keycloak, this leaves tokens valid on the server.
Apache Airflow before 3.4.0 bounded its secrets masker at a nesting depth of five, so secrets registered with the masker appeared in cleartext when nested deeper under non-sensitive keys. Affected output includes task logs and rendered template fields, exposing connection passwords and variable values to anyone who can read tasks.
CVE-2026-111455 in Apache Airflow 3.3.0 before 3.4.0: the async connection test authorizes create at enqueue time but writes the tested connection at commit under that ID. A user with create but not edit permission can therefore overwrite an existing connection. Severity is rated low.
Apache Airflow asset API endpoints returned IDs of Dags scheduled on, producing, or consuming an asset without verifying the caller's read permission on those Dags. An authenticated user with asset read access could thus learn Dag IDs they were not authorized to see.
CVE-2026-111453 in Apache Airflow's API checks Dag edit permission against the ?dag_id query value while the queuedEvents DELETE deletes rows for every Dag the caller can read. An authenticated user with edit rights on one Dag could suppress asset-triggered runs of other Dags they can only read.
Apache Airflow before 3.4.0 only redacted sensitive keys in snake_case form, so secrets under camelCase keys like privateKey, connectionString, webhookUrl and accessKey were not masked in task logs, rendered templates or API responses. The issue was assigned CVE-2026-111452 with low severity.
Apache Airflow versions 3.0.0 before 3.4.0 let task-written XComs bypass the reserved-key check in the Task Execution API. When a user views such a value through the API with deserialization enabled, the API server instantiates airflow.* classes with Dag-controlled arguments.
Apache Airflow 3.2.0 through 3.4.0 did not reject reserved serialization keys in a Dag run's conf. A user who can trigger a Dag with an asynchronous deadline callback could pass a conf that, on a missed deadline, repeatedly crashed the scheduler or rebuilt objects in the Triggerer.
Apache Airflow before 3.4.0 did not mask the apprise provider's `config` connection extra in the connections API, allowing users with connection read permission to retrieve embedded notification-service credentials in cleartext. Fixed in apache-airflow 3.4.0.
Apache Airflow's example_dag_decorator example DAG validated an HTTP-fetched IP using Python's ipaddress module, then inserted it into a shell command. IPv6 zone identifiers with arbitrary characters pass validation, allowing command injection when example DAGs are loaded and that DAG is triggered.
CVE-2026-86829 in Apache Airflow before 3.4.0 allows full deserialization of task-written XCom values inside the scheduler during dependency evaluation. A DAG author thus controls data deserialized in the scheduler, a context Airflow reserves for Deployment Manager-installed code.
Apache Airflow before 3.4.0 masks CLI secrets by argument position in audit logs, so trailing --description on Variable commands, --conn-extra, and inline --password in keycloak-auth-manager commands are logged in cleartext. Rated low severity by the reporter.
Apache Airflow before 3.4.0 rebuilt DeadlineAlert fields via generic deserialization, importing and instantiating the class named in a serialized Dag within the scheduler and API server. A Dag author could construct allow-listed classes in those processes with attacker-controlled arguments.
The Apache Airflow Task SDK built SecretCache keys from team name and secret id using unescaped underscores, allowing distinct (team, secret id) pairs to collide. In multi-team deployments, a DAG author could craft a colliding secret id to read another team's Variable or Connection value.