The Adversarial Robustness Toolbox (ART) through version 1.20.1 allows command‑line argument injection via its Kubeflow component, robustness_evaluation_fgsm_pytorch.py. The script uses an unsafe eval() call to parse the values supplied to the --clip_values and --input_shape arguments, enabling an attacker to execute arbitrary Python code when the script runs. This flaw can result in full system compromise if exploited on a machine where the ART evaluation is performed.

The vulnerability affects the Adversarial Robustness Toolbox, specifically the Kubeflow component robustness_evaluation_fgsm_pytorch.py in releases up to and including 1.20.1. The flaw is tied to the command‑line handling of the --clip_values and --input_shape parameters.

Because the code is executed by the script itself, any attacker who can control the command‑line arguments—such as through a compromised pipeline configuration or a malicious automated script—can achieve arbitrary code execution. The CVSS score of 9.8 indicates a critical severity, and the EPSS score of <1% suggests a low but non‑zero exploitation probability. The vulnerability is not listed in CISA’s KEV catalog. The risk is remote but requires the attacker to provide the vulnerable arguments; therefore, exposure is limited to environments where ART is executed under user‑supplied configurations.