Preface: NVIDIA NeMo is a widely adopted, end-to-end framework for building, customizing, and deploying generative AI models (LLMs) and conversational AI agents. It is primarily used to tailor open-source models—such as Llama, Mistral, and Google Gemma—using proprietary enterprise data.

Ollama, Mistral, and Google Gemma represent a powerful ecosystem for running local, open-weight Large Language Models (LLMs). Ollama acts as the engine to run models, while Mistral and Gemma are two of the most popular, high-performing model families designed to be efficient enough to run on personal computers.

Background: Regarding the restore_from() method, it is a core functionality used to load local checkpoint files with the .nemo extension.

Key Details of restore_from() –

• Purpose: Fully restores a model instance, including its weights and configuration, from a local [.]nemo file for evaluation, inference, or fine-tuning.
• File Structure: A [.]nemo file is an archive (specifically a tar.gz file) containing the model’s weights and a model_config[.]yaml file that defines its architecture.
• Usage: It is called directly from the model’s base class (e.g., ASRModel[.]restore_from(restore_path="path/to/file[.]nemo")).

Vulnerability details: (see below)

CVE-2026-24157 NVIDIA NeMo Framework contains a vulnerability in checkpoint loading where an attacker could cause remote code execution. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure and data tampering.

CVE-2026-24159 NVIDIA NeMo Framework contains a vulnerability where an attacker may cause remote code execution. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure and data tampering.

Official announcement: Please refer to the link for details –

https://nvidia.custhelp.com/app/answers/detail/a_id/5800

Preface: A design limitation has been discovered in the ONNX quantization function of the NVIDIA model optimizer for Windows and Linux. However, confusingly, the ONXY function appears to only work on Windows/RTX (not Linux). What is the actual design limitation?

A sophisticated technical question. The confusion often stems from the fact that while ONNX is the primary deployment format for Windows/RTX, the quantization process (where the vulnerability often lies) frequently occurs on Linux development servers.

Background: Why does the vulnerability affect both Linux and Windows?

Although ONNX is the target format for Windows AI PC applications, the NVIDIA Model Optimizer (ModelOpt) library is cross-platform.

*Linux as the “Factory”: Most developers use powerful Linux servers (with A100/H100 GPUs) to run the ModelOpt quantization scripts. They generate the optimized ONNX model on Linux and then “ship” it to Windows clients. Therefore, the vulnerability exists in the Linux-based conversion tools.

Vulnerability details: NVIDIA Model Optimizer for Windows and Linux contains a vulnerability in the ONNX quantization feature, where a user could cause unsafe deserialization by providing a specially crafted input file. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, data tampering, and information disclosure. (Initial release – March 24, 2026)

Official announcement: Please refer to the link for details – https://nvidia.custhelp.com/app/answers/detail/a_id/5798

Preface: The core design goal of NVIDIA Apex is to achieve mixed precision training, which mainly involves a combination of **16-bit (FP16)** and **32-bit (FP32)**.

Background: In NVIDIA APEX, handling FP16 and FP32 data is primarily managed through the Automatic Mixed Precision (AMP) module. You don’t need to manually cast your data.

• Use FP32 if you are doing scientific simulations that require extreme precision or if your model fails to converge using lower bit-depths.
• Use FP16 for Inference (running a finished model on a phone or server) or when training Large Language Models (LLMs) to save massive amounts of time and electricity.

Current Gold Standard is FP16; balances speed and memory.

Scientific simulations (such as those simulating black holes or aircraft airflow) require extremely high numerical stability to prevent errors from accumulating over time; while LLM training is more like finding a probability distribution of the “general direction,” where speed and model size are more important than accuracy to the 15th decimal place.

Vulnerability details: NVIDIA APEX for Linux contains a vulnerability where an unauthorized attacker could cause a deserialization of untrusted data. This vulnerability affects environments that use PyTorch versions earlier than 2.6. A successful exploit of this vulnerability might lead to code execution, denial of service, escalation of privileges, data tampering, and information disclosure.

Official announcement: Please refer to the link for details –

https://nvidia.custhelp.com/app/answers/detail/a_id/5782

Best Practices:

•Weight Files: Always convert and store your .pth or .bin files as .safetensors.

•API Inputs: Prefer Protocol Buffers (Protobuf) or JSON for real-time requests.

•Integrity Checks: Before deserializing, verify the file’s SHA-256 hash to ensure it hasn’t been tampered with during transit

Preface: NVIDIA InfiniBand switches are based on technology from Mellanox Technologies. Nvidia Spectrum switches are based on technology from Mellanox Technologies. The Spectrum switch ASIC portfolio, originally developed by Mellanox for high-performance Ethernet networking, was rebranded under Nvidia and is now a core component of Nvidia’s networking division. NVIDIA completed the acquisition of Mellanox Technologies, a major supplier of high-performance interconnect technology (switches, NICs), in April 2020 for approximately $7 billion. This strategic move enhanced NVIDIA’s data center networking capabilities, specifically in InfiniBand and Ethernet, to support AI and high-performance computing.

Background: Cumulus Linux is optimized for Ethernet fabrics, while NVOS/Onyx is largely utilized in high-performance InfiniBand environments.

-Key switches supporting NVIDIA Cumulus Linux include:

• Spectrum-4: SN5600, SN5600D, SN5400
• Spectrum-2/3: SN3700, SN3700C, SN4600, SN4700
• Spectrum-1: SN2700, SN2100, SN2745 

Example: Spectrum-4 series (including the SN5600, SN5600D, and SN5400) is a line of physical Ethernet switches (hardware)

Use Cases: Ideal for hyperscale cloud data centers and enterprise AI networks, emphasizing scalability and full customizability.

-NVOS (NVIDIA Onyx) or similar OS typically supports:

Quantum/Quantum-2 InfiniBand: Switches designed for high-performance AI, such as the Quantum-2 series.

Use Cases: Focused on High-Performance Computing (HPC) and large-scale AI training clusters (AI Factories), particularly environments utilizing NVLink for GPU interconnects.

Note: As of early 2026, NVIDIA is focusing on standardizing the management commands (NVUE) across both systems to reduce the complexity of automation workflows when transitioning between different operating systems.

Cumulus Linux (Native Linux): When you SSH in, you land in a standard Debian Linux bash shell. You configure the switch using the NVUE (NVIDIA User Experience) object model via the nv command (e.g., nv set interface swp1…).

Vulnerability Note: The CVEs (CVE-2025-33179/33180) specifically target the NVUE API and CLI engine found in Cumulus Linux 5.x and later.

Vulnerability details:

CVE-2025-33179 NVIDIA Cumulus Linux and NVOS products contain a vulnerability in the NVUE interface, where a low-privileged user could run an unauthorized command. A successful exploit of this vulnerability might lead to escalation of privileges.

CVE-2025-33180 NVIDIA Cumulus Linux and NVOS products contain a vulnerability in the NVUE interface, where a low-privileged user could inject a command. A successful exploit of this vulnerability might lead to escalation of privileges.

Official announcement: Please refer to the link for details – https://nvidia.custhelp.com/app/answers/detail/a_id/5722

Preface: AMD-SB-3042 is a formal advisory for a specific vulnerability, while AMD-SB-3043 is an advisory regarding an analytical tool (SNPeek) used to detect such vulnerabilities.

Background: If AMD zen5 operating AMD SEV-SNP, traditional hypervisor especially VMware or Hyper-V management include encryption will handle by hypervisor embedded in Xen5, so SNPeek collect the traffic is un-encrypted data. Details shown as below:

Limitations of SNPeek If using a tool like SNPeek to intercept traffic on the host side: It can only see data marked as “Shared” (usually for use by the hypervisor to assist with network or disk I/O). Data in Private Memory is always encrypted to SNPeek; the hypervisor cannot read its plaintext content at all. Potential Risk Warning Despite strong hardware encryption, the recently discovered StackWarp (CVE-2025-29943) vulnerability shows that a malicious hypervisor could still influence the execution path of Zen 5 virtual machines by manipulating the CPU’s internal “Stack Engine.” While this doesn’t mean it can directly “read” encrypted memory, it can achieve indirect attacks.

AMD-SB-3043: Analytical Framework (SNPeek)

• Nature: A bulletin regarding a research framework and toolkit for evaluating side-channel risks in Confidential VMs (CVM).
• Core Content: Describes the SNPeek open-source toolkit.
• Function & Purpose:
  • SNPeek is not a single vulnerability but an automated analysis pipeline that uses machine learning to assess how sensitive a CVM application is to side-channel attacks.
  • It helps developers quantify how much information an application might leak when running in encrypted environments like SEV-SNP.
  • It provides configurable attack primitives to help developers locate “weak points” in their code and guides the implementation of mitigations (e.g., oblivious memory access).

Official details and announcement: AMD’s assessment is that all side-channel techniques demonstrated in the paper fall within the category of already known, documented, and out of scope behaviors according to the published SEV/SNP threat model.  AMD has  introduced features on Zen 5 processors—specifically Ciphertext Hiding and PMC Virtualization—that address the ciphertext visibility and HPC based leakage paths highlighted by the researchers.

AMD recommends software developers employ existing best practices, including constant-time algorithms, and avoiding secret-dependent data accesses where appropriate. Please refer to the link for details – https://www.amd.com/en/resources/product-security/bulletin/amd-sb-3043.html

Preface: Artificial intelligence is both harmful and beneficial. Why is it harmful? Fundamentally, it reduces opportunities for low-skilled jobs. Speakers chant slogans like “Smart living, increased productivity.” However, its underlying problems seem difficult to conceal, so you can learn about the latest developments in artificial intelligence from online newspapers and articles. Today, when you seek answers from artificial intelligence, the answers it provides may not be the truth! Why have humans been able to survive and thrive on Earth for thousands of years? The answer is: survival of the fittest.

Background: Megatron-Core and Megatron-LM are open-source tools that are typically used together to train LLMs at scale across GPUs. Megatron-Core expands the capability of Megatron-LM.

NeMo Megatron Bridge is utilized by AI researchers, infrastructure engineers, and developers focused on high-performance training and fine-tuning of large language models (LLMs) and foundation models, particularly those bridging the Hugging Face ecosystem with NVIDIA’s Megatron-Core. NVIDIA H100 GPU introduced support for a new datatype, FP8 (8-bit floating point), enabling higher throughput of matrix multiplies and convolutions. Megatron Bridge uses the NVIDIA TransformerEngine (TE) to leverage speedups from FP8.

While NVIDIA developed Megatron Bridge to facilitate checkpoint conversion between NVIDIA NeMo and other deep learning frameworks, OpenAI utilizes its own internal infrastructure. As of 2026, NVIDIA Megatron Bridge is primarily used by large enterprises, Cloud Service Providers (CSPs), and Sovereign AI initiatives that need to train or deploy open-source models (such as Llama 3, Mistral, or Qwen) at massive scale on NVIDIA hardware.

Vulnerability details:

CVE-2025-33239 NVIDIA Megatron Bridge contains a vulnerability in a data merging tutorial, where malicious input could cause a code injection. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure, and data tampering.

CVE-2025-33240 NVIDIA Megatron Bridge contains a vulnerability in a data shuffling tutorial, where malicious input could cause a code injection. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure, and data tampering.

Official announcement: Please refer to the link for details – https://nvidia.custhelp.com/app/answers/detail/a_id/5781

Preface: NeMo 2.0 is NVIDIA’s major modernization of the NeMo ecosystem.

Two things to remember about NeMo 2.0:

1. NeMo 2.0 is the training & model building framework.

It focuses on:

•               Model architectures (LLMs, ASR, TTS, multimodal)

•               Training pipelines

•               NeMo Run + NeMo-based microservices

•               Distributed GPU/accelerated workflows

2. NeMo Guardrails and NeMo Curator are NOT part of the NeMo 2.0 training stack.

They live adjacent to NeMo 2.0, serving two different lifecycle phases.

Background: NeMo 1.x modules (ASR collections, VAD, etc.) used pickle because they relied heavily on Python multiprocessing and Python objects.

NeMo 2.0 is moving toward language  and framework agnostic formats

Instead of pickle, NeMo 2.0 favors:

•               Safetensors (for weights)

•               JSON / YAML (for metadata)

•               Parquet (for curated datasets)

•               Numpy / torch tensors loaded explicitly

•               HuggingFace compatible formats

These formats are:

•               Safe

•               Portable across hardware and OS

•               Usable by non Python systems

•               Compatible with cloud trust boundaries

NeMo Curator and NeMo Guardrails are designed to avoid pickle entirely

Even though older NeMo components still used pickle internally:

• NeMo Curator does not ingest pickle data
• NeMo Guardrails never used pickle at all
• NeMo 2.0 framework minimizes it or removes it

This aligns with modern security guidance for LLM infrastructure.

Vulnerability details: CVE-2025-33245 NVIDIA NeMo Framework contains a vulnerability where malicious data could cause remote code execution. A successful exploit of this vulnerability might lead to code execution, escalation of privileges, information disclosure, and data tampering.

Official announcement: Please refer to the link for details –

https://nvidia.custhelp.com/app/answers/detail/a_id/5762

Preface: Janet has a small footprint: it fits in environments where 2.5 MB of RAM is considered “plenty”. While Janet does not run on the GPU (it is a CPU-bound language), it is often used as the control/orchestration layer on heterogeneous AI platforms. Because Janet is written in C and compiles to a small binary (roughly 200–300 KB), it is frequently used on ARM Cortex-based systems. While Janet does not run on the GPU (it is a CPU-bound language), it is often used as the control/orchestration layer on heterogeneous AI platforms.

Background: Janet can be used to manage the data pipeline, calling into C/C++ libraries that handle heavy GPU lifting via CUDA. If code allow external scripts or users to submit code dynamically, it can use Janet’s built-in eval-string function. Does it vulnerable to CVE-2026-2242.

My speculation: Using eval-string does expose your Jetson pipeline to CVE‑2026‑2242, because:

CVE‑2026‑2242 is triggered during compilation of Janet code, and eval-string compiles code dynamically. If a malicious user submits a specially-crafted Janet expression that enters the vulnerable path inside:

janetc_if  →  specials[.]c

then Janet may perform an out‑of‑bounds read, which can cause:

• interpreter crash
• denial of service
• undefined behavior inside the Janet process

Even though the CVE requires “local execution,”

-allowing remote users to submit code and then calling eval-string makes that local execution possible.

Therefore, theJetson pipeline becomes exploitable.

Vulnerability details: A vulnerability was determined in janet-lang janet up to 1.40.1. This impacts the function janetc_if of the file src/core/specials.c. Executing a manipulation can lead to out-of-bounds read. The attack needs to be launched locally. The exploit has been publicly disclosed and may be utilized. This patch is called c43e06672cd9dacf2122c99f362120a17c34b391. It is advisable to implement a patch to correct this issue.

Official announcement: Please refer to the link for more details –

https://www.tenable.com/cve/CVE-2026-2242

Preface: Microsoft has launched several products (Semantic Kernel, Microsoft.Extensions.AI, and Azure.OpenAI), which initially caused confusion for developers. Furthermore, the Semantic Kernel is currently being “upgraded” to the new Microsoft Agent Framework, leading some developers to question future support for the Semantic Kernel.

Python, with its rich libraries and large open-source community, remains the “universal language” in artificial intelligence research and data science. LangChain, as the primary alternative, is also based on Python. Python remains dominant in the field of artificial intelligence/machine learning.

Background: Microsoft developed Semantic Kernel as an open-source SDK to bridge conventional programming languages (C#, Python, Java) with advanced LLMs, enabling developers to build enterprise-grade, agentic AI applications. It simplifies orchestrating complex AI workflows, allows swapping models without rewriting code, and ensures secure, compliant integration of AI with existing systems.

Semantic Kernel uses plugins (formerly skills or functions) to extend its capabilities beyond its core prompt engineering functionality and integrate with external services, data sources, and API.

• Semantic Kernel acts as an orchestrator, using the LLM to decide which plugins to use and when, effectively automating complex tasks that involve multiple steps and tools. The LLM determines the necessary sequence of actions to fulfill a user’s request.
• Plugins allow the LLM to interact with real-world applications and data. For example, a plugin could retrieve real-time weather information, search a database, book a flight, or send an email.

The airline and travel industry is beginning to use Microsoft Semantic Kernel to build intelligent, AI-powered applications, particularly for automating customer service and booking processes. Developers are using Semantic Kernel to build dialogue agents that can understand complex booking instructions, such as “book the cheapest flight from Hong Kong to Tokyo”, and handle the booking process independently.

Vulnerability details: CVE-2026-25592 Semantic Kernel is an SDK used to build, orchestrate, and deploy AI agents and multi-agent systems. Prior to 1.70.0, an Arbitrary File Write vulnerability has been identified in Microsoft’s Semantic Kernel .NET SDK, specifically within the SessionsPythonPlugin. The problem has been fixed in Microsoft.SemanticKernel.Core version 1.70.0. As a mitigation, users can create a Function Invocation Filter which checks the arguments being passed to any calls to DownloadFileAsync or UploadFileAsync and ensures the provided localFilePath is allow listed.

Official announcement: Please refer to the link for details –https://nvd.nist.gov/vuln/detail/CVE-2026-25592