Security vulnerabilities and automated fixes for ml security issues
2 posts found
A machine learning application was fetching ONNX model files and chunks over the network without verifying their integrity, creating an opening for model poisoning attacks. The fix adds cryptographic integrity verification at the point where downloaded chunks are reassembled and cached, ensuring models have not been modified in transit or at rest.
A shell injection vulnerability in TensorFlow's DELF dataset download script allowed attackers who controlled the `data_dir` parameter to execute arbitrary shell commands by injecting metacharacters into `os.system()` calls. The fix replaces all four `os.system()` invocations with `subprocess.run()` using argument lists, eliminating shell interpretation entirely. This change closes a high-severity code execution path in production ML infrastructure.