Adversarial machine learning

The Adversarial Robustness Toolbox (ART) through 1.20.1 has a command-line argument injection flaw in its Kubeflow component, robustness_evaluation_fgsm_pytorch.py. The script passes the --clip_values and --input_shape arguments to the unsafe eval() function, so injected Python code runs when eval() is called. An attacker who can control these arguments, for example through pipeline configuration or automated scripts, can achieve arbitrary code execution on the system running the ART evaluation.

The Adversarial Robustness Toolbox (ART) through 1.20.1 contains an insecure deserialization flaw (CWE-502) in its Kubeflow component's model loading. During robustness evaluation, model weights loaded from a file such as model.pt use torch.load() without weights_only=True, which permits Pickle deserialization of arbitrary Python objects. An attacker who uploads a crafted model file to object storage referenced by the pipeline, or who controls the model_id parameter to point to such a file, can achieve remote code execution when the pipeline loads the model.

CVE-2024-5185 affects the EmbedAI application, which is susceptible to security issues enabling Data Poisoning attacks. The flaw stems from a CSRF vulnerability, enabled by the absence of secure session management and weak CORS policies. An attacker who lures a user to a malicious webpage can trick that user into uploading and integrating incorrect data into the application's language model.