CVE-2026-26190: Milvus RCE — Critical — Patch Now
CVE-2026-26190
Patch now: CVE-2026-26190 critical RCE in Milvus <2.5.27 and <2.6.10 grants unauthenticated attackers full admin access with no password, risking data theft and AI compromise. Upgrade to 2.5.27+ or 2.6.10+.
Patch now - CVE-2026-26190 is a critical remote code execution vulnerability in Milvus vector database versions before 2.5.27 and 2.6.10 that grants unauthenticated attackers full administrative control, including data theft, deletion, and server takeover, via exposed management ports with no authentication. Upgrade immediately to the patched versions.
Overview
A critical security vulnerability has been identified in the Milvus vector database, a core component for many generative AI and similarity search applications. This flaw allows attackers to bypass all authentication controls, potentially leading to a complete compromise of the database and its data.
Vulnerability Description
In simple terms, Milvus was shipped with several major security oversights in its default configuration. First, a TCP port used for internal management was unnecessarily exposed. On this port, two critical issues existed:
- A debugging feature was protected by a weak, predictable password that could be easily guessed.
- More severely, the entire operational API-the interface used to insert, query, and manage data-was made available on this port with no authentication whatsoever.
This combination means an attacker who can reach this port can run commands and access data as if they were a fully privileged administrator, without needing a username or password.
Potential Impact
The impact of this vulnerability is severe (CRITICAL, CVSS: 9.8). An unauthenticated remote attacker could:
- Steal, modify, or delete all vector data and metadata stored in the database.
- Compromise AI applications that rely on this data, leading to data leakage, corrupted outputs, or complete service failure.
- Access and manipulate system credentials and configurations, enabling a persistent foothold within the environment.
- Execute arbitrary code on the server via the vulnerable debug endpoint.
Any Milvus instance exposed to a network (including internal networks) is at significant risk.
Remediation and Mitigation
Immediate Action Required:
- Upgrade Immediately: This is the primary fix. Update your Milvus installation to version 2.5.27 or 2.6.10 or later. These versions contain the necessary patches.
- Restrict Network Access: As a critical interim measure until patching, use firewall rules (e.g., AWS Security Groups, Azure NSGs, iptables) to block all external access to TCP port 9091 on your Milvus nodes. Ensure this port is only accessible from strictly necessary, trusted management hosts.
Verification and Best Practices:
- After upgrading, verify that the
/api/v1/endpoints and the/exprdebug endpoint are no longer accessible without valid authentication. - Review your deployment architecture. Milvus management and metrics ports should never be exposed to the public internet. Adhere to the principle of least-privilege network access.
- Regularly subscribe to security announcements for the open-source components in your AI/ML stack.
Note: Simply changing the default etcd.rootPath does not fully mitigate this vulnerability, as the core issue was the complete lack of authentication on the main API. Patching is the only complete solution.
Never miss a critical vulnerability
Get real-time security alerts delivered to your preferred platform.
Am I Affected by CVE-2026-26190?
Pick an ecosystem, paste your installed version, and we'll compare it against the fixed version published on OSV.dev. Browser-only — nothing is sent to a server.
Heuristic comparison only. Always cross-check against the vendor advisory before making patching decisions.
Related Advisories
WordPress MStore API 2.0.6 contains an arbitrary file upload vulnerability that allows unauthenticated attackers to upload malicious files by sending POST requests to the REST API endpoint. Attackers ...
OpenCATS 0.9.4 contains a remote code execution vulnerability that allows unauthenticated attackers to execute arbitrary commands by uploading malicious PHP files disguised as resume attachments. Atta...
Arelle before 2.39.10 contains an unauthenticated remote code execution vulnerability in the /rest/configure REST endpoint that accepts a plugins query parameter and forwards it to the plugin manager ...
marimo is a reactive Python notebook. Prior to 0.23.0, Marimo has a Pre-Auth RCE vulnerability. The terminal WebSocket endpoint /terminal/ws lacks authentication validation, allowing an unauthenticate...