What matters in AI.

Subscribe

AWS will not repair two flaws in its Multi Model Server

AWS says no repair exists or is planned for all versions, and users must move to a tool that it maintains.

On 23 October 2026, Amazon Web Services (AWS) published a bulletin about two flaws in Multi Model Server. This open source tool operates AI models to answer requests. A crafted model archive can write files in an external folder and possibly start code. The management interface has no sign-in step. All versions to 1.1.11 are affected.

What is affected

The first flaw (CVE-2026-115812) is in the extraction of model archives. The second (CVE-2026-115813) is in the model download path. It can make the management interface send requests to unintended network destinations.

What AWS advises

The tool is at end of life. Its source code is read-only since 20 November 2025. AWS gives these steps:

  • Move to a tool that AWS maintains.
  • Where the setup permits, bind the management interface to the loopback address.
  • Do not give external URLs to the model download path.

What is not known

The SageMaker-style setup binds management to all addresses on the same port as inference. There, AWS says to move first and to limit the port to trusted callers. AWS gave no sign of attacks.

Insight: A bad model and an unmaintained server make one path in

In November 2026, researchers found 41 malicious models on Hugging Face, and ModelScan found only 29% of them. This AWS tool loads model archives and no owner repairs it. A poisoned model and an unmaintained loader seem to make one path in, and no vendor will close it. For leaders: record unsupported AI tools as a risk.

Sources

  1. jcr:contentaws.amazon.com
AI MATTER · NEWS · AI MATTER · NEWS ·19 NOV2026

Posted

Tags

More in Security

All Security news