Executive brief
Milvus, an open-source vector database used for AI and machine learning applications, contains a critical security flaw where its management port (9091) is exposed without proper authentication. An attacker can exploit this to gain full control over the database, allowing them to steal sensitive data, delete entire collections, or create new administrative accounts. This could lead to a total compromise of the database system and any integrated AI services.
Technical details
Milvus suffers from two primary authentication bypasses on TCP port 9091. First, the '/expr' debug endpoint uses a predictable default token ('by-dev') to allow arbitrary Go expression evaluation, which can be leveraged for credential theft or arbitrary file writes via access log manipulation. Second, the full REST API ('/api/v1/*') is incorrectly registered on the metrics/management port without any authentication middleware, even when authentication is enabled on primary ports. A remote, unauthenticated attacker can reach this port to perform CRUD operations on data, manage user credentials, or trigger a denial of service. Patches are available in versions 2.5.27 and 2.6.10.
Affected products
- Milvus-io Milvus < 2.5.27, >= 2.6.0, < 2.6.10
Timeline
- 2026-02-11: advisory: GitHub Advisory published
- 2026-02-11: patched: Fixes released in versions 2.5.27 and 2.6.10
- 2026-02-13: disclosed: NVD publication date