Mardi 12 mai 2026

The mamba language model framework thru 2.2.6 is vulnerable to insecure deserialization (CWE-502) when loading pre-trained models from HuggingFace Hub. The MambaLMHeadModel.from_pretrained() method uses torch.load() to load the pytorch_model.bin weight file without enabling the security-restrictive weights_only=True parameter. This allows the deserialization of arbitrary Python objects via the pickle module. An attacker can exploit this by publishing a malicious model repository on HuggingFace Hub. When a victim loads a model from this repository, arbitrary code is executed on the victim's system in the context of the mamba process.

The Ludwig framework thru 0.10.4 is vulnerable to insecure deserialization (CWE-502) in its model serving component. When starting a model server with the ludwig serve command, the framework loads model weight files using torch.load() without enabling the security-restrictive weights_only=True parameter. This default behavior allows the deserialization of arbitrary Python objects via the pickle module. An attacker can exploit this by providing a maliciously crafted PyTorch model file, leading to arbitrary code execution on the system hosting the Ludwig model server.

The Ludwig framework thru 0.10.4 is vulnerable to insecure deserialization (CWE-502) through its predict() method. When a user provides a dataset file path to the predict() method, the framework automatically determines the file format. If the file is a pickle (.pkl) file, it is loaded using pandas.read_pickle() without any validation or security restrictions. This allows the deserialization of arbitrary Python objects via the unsafe pickle module. A remote attacker can exploit this by providing a maliciously crafted pickle file, leading to arbitrary code execution on the system running the Ludwig prediction.

Red Hat OpenShift AI (RHOAI)

The llm CLI tool thru 0.27.1 contains a critical code injection vulnerability via its --functions command-line argument. This argument is intended to allow users to provide custom Python function definitions. However, the tool directly executes the provided code using the unsafe exec() function without any sanitization, sandboxing, or security restrictions. An attacker can exploit this by crafting a malicious llm command with arbitrary Python code in the --functions argument and using social engineering to trick a victim into running it. This leads to arbitrary code execution on the victim's system, potentially granting the attacker full control.

The imgaug library thru 0.4.0 contains an insecure deserialization vulnerability in its BackgroundAugmenter class within the multicore.py module. The class uses Python's pickle module to deserialize data received via a multiprocessing queue in the _augment_images_worker() method without any safety checks. An attacker who can influence the data placed into this queue (e.g., through social engineering, malicious input scripts, or a compromised shared queue) can provide a malicious pickle payload. When deserialized, this payload can execute arbitrary code in the context of the worker process, leading to remote or local code execution depending on the deployment scenario.

Cognee thru v0.4.0 contains a critical remote code execution vulnerability in its notebook cell execution API endpoint. The endpoint is designed to execute arbitrary Python code provided by the user, but it does so using the unsafe exec() function without any sandboxing, validation, or security controls. An attacker can exploit this by sending a specially crafted POST request containing malicious Python code to the execution endpoint. This leads to arbitrary code execution on the Cognee server with the privileges of the server process, allowing complete compromise of the system.

Red Hat OpenShift AI (RHOAI)

The Adversarial Robustness Toolbox (ART) thru 1.20.1 contains a command-line argument injection vulnerability in its Kubeflow component (robustness_evaluation_fgsm_pytorch.py). The script uses the unsafe eval() function to parse string values provided via the --clip_values and --input_shape command-line arguments. This allows an attacker to inject arbitrary Python code into these arguments, which will be executed when eval() is called. The vulnerability can be exploited remotely if an attacker can control these arguments (e.g., through pipeline configuration or automated scripts), leading to arbitrary code execution on the system running the ART evaluation.

The Adversarial Robustness Toolbox (ART) thru 1.20.1 contains an insecure deserialization vulnerability (CWE-502) in its Kubeflow component's model loading functionality. When loading model weights from a file (e.g., model.pt) during robustness evaluation, the code uses torch.load() without the security-restrictive weights_only=True parameter. This allows the deserialization of arbitrary Python objects via the Pickle module. An attacker can exploit this by uploading a maliciously crafted model file to an object storage location referenced by the pipeline, or by controlling the model_id parameter to point to such a file. When the pipeline loads the model, the malicious payload is executed, leading to remote code execution.

Red Hat OpenShift AI (RHOAI)

The Adversarial Robustness Toolbox (ART) thru 1.20.1 contains a remote code execution vulnerability in its Kubeflow component. The robustness evaluation function for PyTorch models uses the unsafe eval() function to dynamically evaluate user-supplied strings for the LossFn and Optimizer parameters without any sanitization or security restrictions. An attacker can exploit this by providing a specially crafted string that contains arbitrary Python code, which will be executed when eval() is called, leading to complete compromise of the system running the ART evaluation.

PySyft (Syft Datasite/Server) versions 0.9.5 and earlier are vulnerable to remote code execution due to insufficient validation and sandboxing of user-submitted code. The system allows low-privileged users to submit Python functions (via @sy.syft_function()) for remote execution on the server. While a code approval mechanism exists, the submitted code undergoes no security checks for dangerous operations (e.g., file access, command execution). Once approved, the code is executed within the server process using exec() and eval() functions without proper isolation. A remote attacker can leverage this to execute arbitrary Python code on the server, leading to complete compromise of the server environment.

Nebuly optimate

The _load_model() function in the neural_magic_training.py script of the optimate project in commit a6d302f912b481c94370811af6b11402f51d377f (2024-07-21) allows arbitrary code execution. When a user supplies a directory path via the --model command-line argument, the function reads a module.py file from that directory and executes its contents directly using Python's exec() function. This design does not validate or sanitize the file's content, allowing an attacker who controls the input directory to execute arbitrary Python code in the context of the process running the script.

The torch-checkpoint-shrink.py script in the ml-engineering project in commit 0099885db36a8f06556efe1faf552518852cb1e0 (2025-20-27) contains an insecure deserialization vulnerability (CWE-502). The script uses torch.load() to process PyTorch checkpoint files (.pt) without enabling the security-restrictive weights_only=True parameter. This oversight allows the deserialization of arbitrary Python objects via the pickle module. A remote attacker can exploit this by providing a maliciously crafted checkpoint file, leading to arbitrary code execution in the context of the user running the script.