Executive brief
Deep Java Library (DJL) is Amazon's open-source framework for building deep learning applications in Java. An integer overflow vulnerability in tensor buffer validation allows attackers to bypass size checks and trigger out-of-bounds memory reads. An unauthenticated remote attacker could exploit this to leak sensitive data from adjacent process memory or crash the application, affecting any system using vulnerable versions of DJL.
Technical details
CVE-2026-85228 is an integer overflow vulnerability in the tensor buffer validation component of AWS Deep Java Library. The vulnerability exists in the code that validates tensor payload shapes and computes buffer sizes. An attacker can craft a malicious tensor payload with a shape whose computed byte size exceeds the 32-bit signed integer range, causing integer wrapping and allowing an undersized buffer to pass validation. Subsequent tensor operations attempt to read from this insufficiently-sized buffer, resulting in out-of-bounds memory access. The vulnerability is remotely exploitable without authentication and affects all platforms running DJL versions 0.13.0 through 0.36.0. Exploitation can lead to information disclosure from adjacent process memory or denial of service. A patch is available in version 0.37.0 and later.
Affected products
- AWS Deep Java Library (DJL) >=0.13.0, <=0.36.0
Timeline
- 2026-09-10: disclosed: CVE-2026-85228 disclosed by AWS Security Bulletin 2026-106-AWS
- 2026-09-10: patched: Fix released in DJL version 0.37.0 and later