Executive brief
MLflow is an open-source platform used by teams to manage machine learning workflows and AI models. A security flaw in its AI Gateway component allows attackers to trick the system into revealing sensitive server-side credentials, such as AWS access keys. If exploited, this could allow an attacker to steal cloud credentials, potentially leading to unauthorized data access or the ability to tamper with machine learning models.
Technical details
A vulnerability in MLflow's AI Gateway component (prior to version 3.11.0) stems from the improper handling of the `api_key` field in gateway secrets. The system resolves `$ENV_VAR` references within this field against the server's environment variables at runtime. An attacker can configure a gateway secret to reference sensitive variables (e.g., `AWS_SECRET_ACCESS_KEY`) and set the `api_base` to an external endpoint they control. When the gateway attempts to authenticate to the upstream provider, it sends the resolved secret in the authentication header to the attacker's server. This can be exploited by low-privileged users in basic-auth environments or unauthenticated users in default deployments. The fix in version 3.11.0 gates this resolution behavior behind a new configuration flag.
Affected products
- MLflow MLflow < 3.11.0
- Red Hat Red Hat OpenShift AI (RHOAI)
Timeline
- 2026-06-03: disclosed
- 2026-06-03: advisory
- 2026-07-10: patched: GitHub Advisory updated with patch details
References
- https://github.com/mlflow/mlflow/commit/4a3f2f720cb4f058c9e0c5b883e0acc9ab64a7f3
- https://huntr.com/bounties/f8e591a0-0f19-4910-b82e-16c9956f2233
- https://access.redhat.com/security/cve/CVE-2026-4035
- https://bugzilla.redhat.com/show_bug.cgi?id=2484318
- https://security.access.redhat.com/data/csaf/v2/vex/2026/cve-2026-4035.json