Security vulnerabilities and automated fixes for machine learning security issues
2 posts found
A high-severity untrusted deserialization vulnerability was discovered in `TransferLearningTF.ipynb`, a transfer learning tutorial notebook that loads VGG16 model weights from the internet without verifying their integrity. Because Keras relies on Python's pickle-based serialization format under the hood, a tampered or substituted weights file could execute arbitrary code with the full privileges of the notebook user. The fix adds a SHA-256 checksum verification step immediately after the weight
A critical command injection vulnerability was discovered in DeepSpeed's `data_analyzer.py`, where an `os.system()` call directly interpolated an unsanitized file path variable into a shell command string. An attacker who could influence dataset configuration or file paths could execute arbitrary shell commands on the host machine. The fix replaces the dangerous shell invocation with safe, Python-native file operations that never touch a shell interpreter.