Skip to content
deepjavalibrary/djl·Indexed by this site · Originally published· Event →· Its place in the scan record →

Deep Java Library model archive metadata vulnerability can lead to arbitrary local file reads, fixed in 0.40.0

Arbitrary local file read via unvalidated resource location in model archive metadata in Deep Java Library

AI summary

Deep Java Library (DJL) released security advisory GHSA-xf3w-qp3m-fp62 (CVE-2026-129266): the resource location specified by model archive metadata does not validate the scheme, and this can be used to read arbitrary local files. After a crafted model archive writes a local path into the metadata, DJL will open and read that file and return its contents to the caller as model data; a remote unauthenticated attacker able to publish or provide a model archive to affected users can read any file accessible to the library process. Affected versions are ai.djl:api 0.7.0 through 0.39.0, and it has been fixed in 0.40.0. Mitigations include loading model archives only from trusted sources, and restricting network access to the management and inference APIs when running djl-serving or the djl-inference Deep Learning Container that bundles it. DJL thanks Oguzhan Akkaya for participating in this issue through the coordinated vulnerability disclosure process.

Read originalgithub.com