Path Traversal in Keras 3 Model Saving Library

Executive Summary

This vulnerability is a path traversal issue in the keras-team/keras version 3.14.0, specifically in the DiskIOStore.make method used for saving and loading Keras models.

It occurs because user-provided layer names are used to build directory paths without properly sanitizing for parent directory components like ".." sequences.

Although forward slashes are restricted in layer names, directory traversal sequences are not, allowing an attacker to craft a malicious model that escapes the intended temporary directory.

This can lead to unauthorized file system operations such as creating directories or writing files in arbitrary locations.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability can allow an attacker to perform unauthorized file system operations outside the intended directory.

Potential impacts include creating or modifying files and directories in arbitrary locations, which could lead to data corruption, unauthorized data access, or disruption of system operations.

Because the vulnerability requires user interaction (loading or saving a malicious model), it may be exploited in environments where untrusted models are used.

Useful 0 Not Useful 0