AWS fixes a bug in an AutoGluon tool that can write files
Only an optional tool that gets public datasets has the bug, and AutoGluon version 1.6.0 fixes it.
On 18 November 2026, AWS gave information about CVE-2026-130638, a bug in AutoGluon MultiModal, an open source machine learning framework that AWS maintains. A tool that prepares object detection datasets downloaded archives with HTTP and did not check the files in them. An attacker on the network path can send a different archive and write files in the incorrect folder.
How it works
The tool is an optional command line utility that gets public benchmark datasets. HTTP has no encryption, and an attacker between the user and the server can send a different archive. The files then go on the disk with the privileges of the user who operates the tool.
Who is exposed
AWS says the training and inference APIs do not have the bug, also object detection with MultiModalPredictor. Versions before 1.6.0 have the bug. The bulletin gives no count of users and no test results.
What is not known
The bulletin does not tell if anyone used the bug in an attack. If you cannot upgrade, do not operate the prepare_detection_dataset tool. Get the datasets from their official sources with HTTPS.
Insight: The weak point is the small tool around the AI core
AWS says the AutoGluon training tools are safe and an optional dataset tool is not. Check Point found that a ChatGPT HTML preview, not the model, broke Lockdown Mode. On 17 November 2026, a Microsoft agent package operated a stealer at each Python start, also when no code used it. Attackers seem to go for side tools that get less inspection than the core. Count the helper tools in your AI stack.
Sources
- CVE-2026-130638Path traversal in the object detection dataset preparation tool in AutoGluon MultiModal
Posted