Official Updated 08/11/2025 06:16 AM

Preface: GPT-4 offers several key benefits, including improved accuracy, longer context handling, and the ability to process both text and image inputs. It also exhibits stronger guardrails, leading to more reliable and ethical outputs. Additionally, GPT-4 excels in various tasks like professional and academic benchmarks, creative writing, and adapting to user needs.

Background: The Megatron-LM codebase is a framework for training large, powerful transformer language models at scale, developed by NVIDIA. It focuses on efficient, model-parallel (tensor and pipeline) and multi-node pre-training of transformer-based models like GPT, BERT, and T5 using mixed precision

Megatron-LM codebase efficiently trains models from 2B to 462B parameters across thousands of GPUs, achieving up to 47% Model FLOP Utilization (MFU) on H100 clusters.

GPT-4, the latest iteration in OpenAI’s Generative Pre-trained Transformer series, significantly scales up the parameter count compared to its predecessors. While GPT-2 had 1.5 billion parameters and GPT-3 boasted 175 billion, GPT-4 is estimated to have a staggering 1.76 trillion parameters.

The Megatron-LM codebase has successfully benchmarked the training of a 462B parameter model using 6144 H100 GPUs, achieving up to 47% Model FLOP Utilization (MFU).

While this demonstrates the capability of the Megatron-LM framework to train very large models on H100 clusters, the exact number of H100 GPUs used to train GPT-4 is not publicly disclosed. GPT-4 was developed by OpenAI, and they have not released the specific hardware configurations used for its training.

Vulnerability details:

CVE-2025-23305       NVIDIA Megatron-LM for all platforms contains a vulnerability in the tools component, where an attacker may exploit a code injection issue. A successful exploit of this vulnerability may lead to code execution, escalation of privileges, information disclosure, and data tampering.

CVE-2025-23306       NVIDIA Megatron-LM for all platforms contains a vulnerability in the megatron/training/

arguments.py component where an attacker could cause a code injection issue by providing a malicious input. A successful exploit of this vulnerability may lead to code execution, escalation of privileges, information disclosure, and data tampering.

Official announcement: For more information, please refer to the link

https://nvidia.custhelp.com/app/answers/detail/a_id/5685

Official Updated 08/11/2025 06:15 AM

Preface: While the Bible doesn’t specifically mention artificial intelligence, it reminds us that human knowledge and capabilities will increase dramatically in the last days (Daniel 12:4). Building and training neural networks is a cornerstone of modern artificial intelligence, enabling breakthroughs in fields such as computer vision, natural language processing, and robotics.

Background: NVIDIA Merlin Transformers4Rec is a Python library designed for building sequential and session-based recommender systems, leveraging the power of Transformer architectures, particularly for use with PyTorch. It is part of the broader NVIDIA Merlin ecosystem, which provides end-to-end GPU-accelerated solutions for recommender systems.

Transformers4Rec is pre-installed in the merlin-pytorch container that is available from the NVIDIA GPU Cloud (NGC) catalog.

NVIDIA Merlin PyTorch container, available on NVIDIA NGC (NVIDIA GPU Cloud), includes the necessary components for GPU acceleration, including the CUDA Toolkit.

The Merlin PyTorch container allows users to do preprocessing and feature engineering with NVTabular, and then train a deep-learning based recommender system model with PyTorch, and serve the trained model on Triton Inference Server.

Ref: NVTabular and RAPIDS (cuDF/cuML) for preprocessing and feature engineering.

Vulnerability details: NVIDIA Merlin Transformers4Rec for all platforms contains a vulnerability in a python dependency, where an attacker could cause a code injection issue. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure, and data tampering.

Official announcement: Please see the link for details

https://nvidia.custhelp.com/app/answers/detail/a_id/5683

Official Updated 08/11/2025 06:15 AM

Preface: WebDataset is a PyTorch IterableDataset implementation designed for efficient access to large datasets stored in POSIX tar archives. It focuses on sequential/streaming data access, which offers substantial performance advantages in environments where local storage is limited or I/O bottlenecks are a concern. WebDataset is particularly well-suited for very large-scale training, as it minimizes the need for local storage and allows for efficient data loading from various sources, including cloud storage.

Background: NVIDIA WebDataset refers to the integration of WebDataset with NVIDIA technologies like DALI or NeMo, rather than a separate NVIDIA-specific installation. Installing WebDataset itself is straightforward, as it is a Python library.

• DALI is a portable, open-source software library for decoding and augmenting images, videos, and speech to accelerate deep learning applications.

DALI itself doesn’t extract .tar files directly — instead, it processes data streamed from tarballs via WebDataset or other loaders.

• NVIDIA NeMo is a framework for building and deploying generative AI models, particularly those used in conversational AI like speech recognition and natural language processing.

It may extract or stream data depending on the configuration, but tarball handling is abstracted behind the data pipeline.

Vulnerability details: CVE-2025-23294 – NVIDIA WebDataset for all platforms contains a vulnerability where an attacker could execute arbitrary code with elevated permissions. A successful exploit of this vulnerability might lead to escalation of privileges, data tampering, information disclosure, and denial of service.

Official announcement: Please see the link for details

https://nvidia.custhelp.com/app/answers/detail/a_id/5658

Preface: Deep learning in AI generally learns much faster than humans in specific, narrow tasks, especially those involving large datasets and complex computations. However, humans still excel at general intelligence, creative problem-solving, and learning with limited data.

Perhaps, AI does not have this advantage yet!

Background: Keras 3.0 is a major rewrite of the Keras deep learning API, designed to provide a unified and flexible platform for building and deploying deep learning models. Its most significant feature is its multi-backend architecture, allowing users to run Keras workflows on top of various popular deep learning frameworks.

TensorFlow is a comprehensive, low-level machine learning framework capable of building and training models directly. However, Keras plays a crucial role as its official high-level API, providing several benefits that make deep learning development significantly easier and more efficient within the TensorFlow ecosystem.

Keras 3.0 does it work in lambda layer? Yes, the Lambda layer continues to be available and functional in Keras 3.0. In machine learning, specifically within the context of deep learning frameworks like Keras or TensorFlow, a Lambda layer is a type of layer that allows you to wrap arbitrary expressions or functions as a layer in your neural network model.

Vulnerability details: A safe mode bypass vulnerability in the `Model.load_model` method in Keras versions 3.0.0 through 3.10.0 allows an attacker to achieve arbitrary code execution by convincing a user to load a specially crafted `.keras` model archive.

Official announcement: Please see the link for details

https://www.tenable.com/cve/CVE-2025-8747

Preface: Nvidia’s security advisories released on August 4, 2025 (e.g., CVE-2025-23318, CVE-2025-23319) are specifically related to the Python backend. The Triton backend for Python. The goal of Python backend is to let you serve models written in Python by Triton Inference Server without having to write any C++ code.

Background: NVIDIA Triton Inference Server is an open-source inference serving software that streamlines the deployment and execution of AI models from various deep learning and machine learning frameworks. It achieves this flexibility through a modular system of backends. 

Each backend within Triton is responsible for executing models from a specific framework. When an inference request arrives for a particular model, Triton automatically routes the request to the necessary backend for execution. 

Key backend frameworks supported by Triton include:

• TensorRT: NVIDIA’s high-performance deep learning inference optimizer and runtime.
• TensorFlow: A popular open-source machine learning framework.
• PyTorch: Another widely used open-source machine learning library.
• ONNX: An open standard for representing machine learning models.
• OpenVINO: Intel’s toolkit for optimizing and deploying AI inference.
• Python: A versatile backend that can execute models written directly in Python and also serves as a dependency for other backends.
• RAPIDS FIL (Forest Inference Library): For efficient inference of tree models (e.g., XGBoost, LightGBM, Scikit-Learn).

This modular backend architecture allows Triton to provide a unified serving solution for a wide range of AI models, regardless of the framework they were trained in.

Vulnerability details:

CVE-2025-23318: NVIDIA Triton Inference Server for Windows and Linux contains a vulnerability in the Python backend, where an attacker could cause an out-of-bounds write. A successful exploit of this vulnerability might lead to code execution, denial of service, data tampering, and information disclosure.

CVE-2025-23319: NVIDIA Triton Inference Server for Windows and Linux contains a vulnerability in the Python backend, where an attacker could cause an out-of-bounds write by sending a request. A successful exploit of this vulnerability might lead to remote code execution, denial of service, data tampering, or information disclosure.

Official announcement: Please see the link for details –

https://nvidia.custhelp.com/app/answers/detail/a_id/5687

Preface: The NVIDIA Triton Inference Server API supports both HTTP/REST and GRPC protocols. These protocols allow clients to communicate with the Triton server for various tasks such as model inferencing, checking server and model health, and managing model metadata and statistics.

Background: NVIDIA Triton™ Inference Server, part of the NVIDIA AI platform and available with NVIDIA AI Enterprise, is open-source software that standardizes AI model deployment and execution across every workload.

The Asynchronous Server Gateway Interface (ASGI) is a calling convention for web servers to forward requests to asynchronous-capable Python frameworks, and applications. It is built as a successor to the Web Server Gateway Interface (WSGI).

NVIDIA Triton Inference Server integrates a built-in web server to expose its functionality and allow clients to interact with it. This web server is fundamental to how Triton operates and provides access to its inference capabilities on both Windows and Linux environments.

Vulnerability details: CVE-2025-23310 – NVIDIA Triton Inference Server for Windows and Linux contains a vulnerability where an attacker could cause stack buffer overflow by specially crafted inputs. A successful exploit of this vulnerability might lead to remote code execution, denial of service, information disclosure, and data tampering.

Official announcement: Please refer to the link for details –

https://nvidia.custhelp.com/app/answers/detail/a_id/5687

Preface: Memory Allocation Flow:

1. User-space request (e.g., CUDA malloc or OpenGL buffer allocation).
2. Driver calls memmgrCreateHeap_IMPL() to create a memory heap.
3. Heap uses pmaAllocatePages() to get physical memory.
4. Virtual address space is mapped using UVM or MMU walker.
5. Memory is returned to user-space or GPU context.

Background:

An OS-agnostic binary is a compiled program designed to run on multiple operating systems without requiring separate builds for each. This means the binary file can be executed on different OS platforms without modification, achieving a level of portability that’s not common with traditional compiled software.

The core loadable module within the NVIDIA vGPU software package is the NVIDIA kernel driver, specifically named nvidia[.]ko. This module facilitates communication between the guest virtual machine (VM) and the physical NVIDIA GPU. It’s split into two main components: an OS-agnostic binary and a kernel interface layer. The OS-agnostic component, for example, nv-kernel[.]o_binary for the nvidia[.]ko module, is provided as a pre-built binary to save time during installation. The kernel interface layer is specific to the Linux kernel version and configuration.

Vulnerability details:

CVE-2025-23285: NVIDIA vGPU software contains a vulnerability in the Virtual GPU Manager, where a malicious guest could cause a stack buffer overflow. A successful exploit of this vulnerability might lead to code execution, denial of service, information disclosure, or data tampering.

CVE2025-23283: NVIDIA vGPU software for Linux-style hypervisors contains a vulnerability in the Virtual GPU Manager, where a malicious guest could cause stack buffer overflow. A successful exploit of this vulnerability might lead to code execution, denial of service, escalation of privileges, information disclosure, or data tampering.

Official announcement: Please see the url for details –

https://nvidia.custhelp.com/app/answers/detail/a_id/5670

Preface: “LeftoverLocals” allows recovery of data from GPU local memory created by other processes on Apple, Qualcomm, AMD, and Imagination GPUs. LeftoverLocals affects the security posture of the entire GPU application, especially LLM and machine learning models running on affected GPU platforms. NVD published on January 16, 2024. So far, AMD appears to be the only company actively taking remediation measures.

Background: Researchers from Trail of Bits reported a potential vulnerability, titled “LeftoverLocals” article to public on 16th January 2024. The corrective action was taken by AMD in following schedule.

2025-07-18: Updated the Mitigation section for AMD Radeon Graphics

2025-06-23: Updated the Mitigation section for Data Center Graphics, AMD Radeon Graphics, and revised Client Processors table

2025-04-07: Updated the Mitigation section for Data Center Graphics, AMD Radeon Graphics, and Client Processors

2025-02-11: Updated the Mitigation section – Data Center Graphics

2025-01-15: Mitigation section has been updated and AMD Ryzen™ AI 300 Series Processor (Formerly codenamed) “Strix Point” FP8 has been added to the Client Processors list

2024-11-07: Mitigation has been updated for MI300 and MI300A

Updated driver version from 24.x.y to 25.x.y

2024-10-30: Updated mitigation targets

2024-08-02: Updated AMD Software: Adrenalin Edition and PRO Edition versions.

Removed: AMD Ryzen™ 3000 Series Processors with Radeon™ Graphics (Not affected)

Added: AMD Ryzen™ 8000 Series Processors with Radeon™ Graphics and AMD Ryzen™ 7030 Series Processors with Radeon™ Graphics

2024-07-30: Updated the Mitigation section of AMD RadeonTM Graphics and Client processors product tables

Updated Data Center Graphics Inter-VM and Bare Metal/Intra-VM Mitigation product tables

Updated mitigation section month for driver update rollout

2024-05-07: Added Vega products and Mitigation section with Product tables

2024-01-26: Updated Graphics and Data Center Graphics products

2024-01-16: Initial publication

Vulnerability details: CVE-2023-4969: A GPU kernel can read sensitive data from another GPU kernel (even from another user or app) through an optimized GPU memory region called _local memory_ on various architectures.

Official announcement: Please refer to the official link for details – https://www.amd.com/en/resources/product-security/bulletin/amd-sb-6010.html

Remark: In step 5, CU2 is written incorrectly. The correct word should be CU.