While working on moon-dev-ai-agents project, I found a vulnerability in Keras (CVE-2025-12060). The keras.utils.get_file API is vulnerable to path traversal when extract=True is used with malicious tar archives. Because it relies on tarfile.extractall without the filter="data" option, attackers can craft tar files with special symlinks that allow writing files outside the intended directory, leading to arbitrary file write on the filesystem.
CVE Link
CVE Report