Overview

    Advance Cybersecurity and Secure AI Workloads With AI

    As AI puts more power in attackers’ hands, defenders need to move faster. NVIDIA brings accelerated computing and AI to every layer of cybersecurity—from earlier threat detection and faster response to protecting sensitive data and IP with zero-trust infrastructure. Its AI frameworks, reference designs, architectures, and best practices give organizations the foundation to build secure, scalable AI data centers.

    Build and Run Safe AI Agents

    Introducing NVIDIA OpenShell Reference Architecture, a full-stack industrial-scale infrastructure enabling safe agents across AI factories, enterprise infrastructure, and the edge.

    Read the Announcement

    Operate Secure AI Agents at Scale

    Together, NVIDIA OpenShell secure runtime and NVIDIA Vera CPUs govern AI agents more efficiently, at industrial scale.

    Read the Technical Blog

    Benefits

    Advance AI Safety and Security With NVIDIA Technology

    Build Agents That Operate Safely

    Open models and tools to filter unsafe inputs and outputs, apply programmable safety controls at runtime without retraining, and proactively scan agents for vulnerabilities—so risks are caught and addressed before reaching production.

    Enforce In-Silicon Security at AI Agent Speed

    NVIDIA Sentry and NVIDIA BlueField deliver out-of-band, in-silicon security for AI agents—providing continuous monitoring, trusted telemetry, and millisecond-scale quarantine.

    Performant Confidential Computing

    Uncover insights with confidence that private data and models remain secure, compliant, and uncompromised—even when sharing datasets or infrastructure with competing or untrusted parties.

    Let Agents Work Without Compromising Security

    A secure runtime enforces declarative policies around file access, network egress, and credentials—so autonomous agents can operate without putting data or infrastructure at risk.

    Technology

    Explore Our Cybersecurity AI Technologies

    NVIDIA’s technologies help enterprises defend their systems and cybersecurity providers strengthen their solutions by leveraging AI and accelerated computing. AI enables efficient processing of large volumes of real-time data, accelerating threat detection and risk identification. NVIDIA also helps organizations secure AI infrastructure, data, and models with reference designs, networking, and confidential platforms.

    Software

    NVIDIA OpenShell

    • Build and deploy autonomous, self-evolving agents more safely with an open source runtime.
    • Govern agent behavior with sandboxed execution and policy enforcement to safely scale agentic systems for your enterprise.
    Learn More About OpenShell
    Hardware and Software

    NVIDIA Confidential Computing

    • Protect sensitive data and model IP with hardware-based trusted execution environments, enabling verifiable AI deployment anywhere.
    • Uncover AI insights while ensuring models and data are secure, compliant, and uncompromised.
    Learn More About Confidential Computing
    Hardware and Software

    NVIDIA Sentry on BlueField

    • Enforce an independent security boundary from BlueField-4’s hardware-isolated, host-independent domain—outside the agent and host software—as part of NVIDIA Open Agent Safety Platform.
    • Continuously monitor agent behavior with NVIDIA Sentry to identify suspicious behavior and policy deviations in real time.
    • Respond to detected threats and quarantine agents in milliseconds using BlueField-4’s in-silicon infrastructure controls.
    Learn More About NVIDIA Sentry Learn More About BlueField
    Models

    NVIDIA Nemotron

    • Power cyber-defense agents with highly efficient, open, multimodal models built for long-running autonomous agents.
    • A family of specialized open models for every layer of agent work—frontier reasoning for planning, fast and efficient execution models for high-volume tool calls, and dedicated safety models that detect unsafe content across text and images with high accuracy.
    • Nemotron™ Safety models detect and filter unsafe inputs and outputs across agentic pipelines, keeping multi-agent cyber systems reliable and aligned.
    Learn More About Nemotron
    Blueprints

    NVIDIA NemoClaw

    • Open reference stack for specialized agents built with Nemotron open models, a tuned open harness, optimized for accuracy, efficiency, and cost.
    • Build specialized agents you can customize, improve, and run anywhere—using open models, harness, and runtime.
    • Powers specialized agents that work alongside people and other AI agents—from clinical teams and autonomous engineers to enterprise operations and personal assistants.
    Start Building
    Software

    NVIDIA NeMo

    • Specialize cybersecurity AI agents for continuous threat hunting, red team simulation, and autonomous cyber defense.
    • Orchestrate enterprise-grade guardrails that keep agentic AI applications safe, reliable, and aligned.
    • Scan agents for vulnerabilities—prompt injection, jailbreaks, data leakage—before they reach production.
    Learn More About NeMo

    NVIDIA OpenShell Reference Architecture for Safe AI Agents

    A full-stack, open architecture powered by the NVIDIA OpenShell secure runtime, Vera CPUs, Vera Rubin systems, and NVIDIA Sentry on BlueField-4.

    Learn More

    Use Cases

    How Cybersecurity AI Is Being Used

    See how NVIDIA AI and accelerated computing support industry use cases, and jump-start your cybersecurity AI development with reference examples.

    Vulnerability Analysis for Container Security

    Addressing software security issues is challenging and time-consuming, but security AI agents can improve vulnerability defense while reducing the burden on security teams. Using NVIDIA NIM and NVIDIA Morpheus, this event-driven RAG application dramatically decreases common vulnerabilities and exposure (CVE) analysis and remediation time from days to just seconds.

    Try Now

    Anomaly Detection

    Enterprises are faced with an incredibly vast network of data to protect. NVIDIA Morpheus enables digital fingerprinting through monitoring of every user, service, account, and machine across the enterprise data center to determine when suspicious interactions occur. Combined with NVIDIA GPU and DPU accelerators and NVIDIA DOCA telemetry in NVIDIA-Certified servers, this brings a new level of security to data centers.

    Learn More About Anomaly Detection

    Data Leakage Prevention

    Traditional ways of detecting leaked sensitive data rely on static, rules-based models, which are limited by the quality of training data. Instead, NVIDIA Morpheus examines raw packet information as it’s generated for potential leakage. The DOCA telemetry agent, residing on the NVIDIA BlueField DPU, pipes raw packets directly to Morpheus. A natural language processing (NLP) model determines if sensitive information—such as passwords and private keys—is being leaked in the packet. Packets are flagged instantaneously, and a recommended action is routed back to DOCA for policy enforcement. These real-time alerts are delivered to the operator so remediation can begin immediately on compromised data.

    Watch a Data Leakage Protection Example

    Spear Phishing Detection

    Spear phishing, one of the largest and most costly cyber threats, uses targeted and convincing emails. It is difficult to defend against due to a lack of training data. NVIDIA Morpheus provides an NLP model that has been trained using synthetic emails generated by NVIDIA NeMo to identify spear phishing attempts. With this, detection of spear phishing emails has improved by 20%—with less than a day of training.

    Learn More About Spear Phishing Detection

    Zero-Trust Architecture

    The traditional perimeter-only security model is insufficient in today’s world of expanding threats and zero-day exploits. NVIDIA BlueField DPUs enable distributed security functions, such as firewalls, encryption, microsegmentation, intrusion detection/prevention, and application inspection, to run on every server, enabling a zero-trust security stance in the data center. Paired with NVIDIA Spectrum-X™ Ethernet networking, NVIDIA BlueField DPUs create a zero-trust, multi-tenant fabric that secures and isolates AI workloads at scale.

    Learn More About Zero Trust

    Securing Enterprise AI Agents

    Securing AI agents is a nonnegotiable requirement for production deployments, as it provides the critical layer of defense against threats. Unlike conventional software that follows fixed logic, AI agents reason and execute code dynamically, making them susceptible to indirect prompt injection and goal hijacking where malicious instructions are hidden in external data (like emails or PDFs). By isolating an agent’s execution with the NVIDIA OpenShell™ secure runtime, organizations enforce strict zero-trust boundaries and guardrails to protect their data.

    Try NVIDIA OpenShell
    Explore All Use Cases

    Ecosystem

    Leading Adopters Across All Industries

    Industry leaders from across the AI ecosystem are joining NVIDIA to strengthen AI safety for every industry across the full stack of infrastructure, software, models, and robotics.

    Agentic AI — Logo Crawl

    NVIDIA AI Security Research

    Explore NVIDIA research on securing AI systems and strengthening cyber defense. Through open research and contributions to the Open Secure AI Alliance, NVIDIA shares methods and technical artifacts that help security teams validate findings and build stronger defenses.

    Explore AI Security Research

    Next Steps

    Ready to Get Started?

    Find everything you need to start developing your cybersecurity AI application, including the latest documentation, tutorials, technical blogs, and more.

    Start Developing

    Get in Touch

    Talk to an NVIDIA product specialist about moving from pilot to production with the security, API stability, and support of NVIDIA AI Enterprise.

    Contact Us

    Stay Up to Date on NVIDIA AI News

    Sign up for the latest news, updates, and more from NVIDIA.

    Stay Informed

    FAQs

    The architecture combines the NVIDIA OpenShell secure runtime with NVIDIA Vera CPUs for agent harnesses and tools, NVIDIA Vera Rubin for inference, and NVIDIA Sentry on BlueField-4 for independent monitoring and policy enforcement. OpenShell also runs on other CPU systems; Vera and BlueField are not required to use it.

    Technical documentation for NVIDIA OpenShell open source runtime, which is used to build and deploy autonomous, self-evolving agents more safely, is available here.

    NVIDIA OpenShell secure runtime isolates an agent’s execution, creating strict zero-trust boundaries and guardrails to protect against threats like indirect prompt injection and goal hijacking, where malicious instructions are hidden in external data.

    NVIDIA BlueField is a purpose-built infrastructure processor that accelerates networking, storage, and security. Combined with NVIDIA DOCA, BlueField provides a hardware-isolated, host-independent security domain that powers NVIDIA Sentry—delivering continuous out-of-band agent monitoring, in-silicon policy enforcement, and millisecond-scale quarantine as part of NVIDIA Open Agent Safety Platform. NVIDIA BlueField is a purpose-built infrastructure processor that accelerates networking, storage, and cybersecurity. Combined with NVIDIA DOCA, it enables zero-trust security across every system, using real-time workload visibility to detect and stop runtime attacks.

    NVIDIA DOCA Argus provides deep visibility into the state and behavior of AI workloads to detect runtime threats at AI speed. Using real-time memory analysis, it builds situational awareness for each workload and triggers alerts when behavior deviates from expected patterns. DOCA Argus is integrated with leading enterprise XDR security platforms.

    NVIDIA DOCA Vault is a security microservice that runs on NVIDIA BlueField to protect sensitive data with in-silicon security, enforce zero-trust access, and reduce the risk of data exposure at the speed of AI agents.

    Confidential computing is a technology that enables enterprises to securely run AI on sensitive data by protecting models and data in use with hardware-rooted Trusted Execution Environments (TEEs)—providing cryptographic verification that data is protected from insider and infrastructure threats.

    Nemotron models are highly efficient, open, multimodal models that can be used to power long-running autonomous cyber-defense agents.

    NVIDIA offers open models and tools that contribute to agent safety, including Nemotron Safety models, LLMs to classify and filter unsafe inputs and outputs in your agent pipeline, NeMo Guardrails for adding programmable safety and topic controls to LLM applications, and Garak for probing LLMs and agents for weaknesses like prompt injection, jailbreaks, and data leakage.

     NVIDIA NemoClaw™ is a collection of open blueprints for specialized agents, optimized for best quality at lowest cost, that operate more securely using NVIDIA OpenShell, an open source runtime for agent execution. Available on GitHub, OpenShell helps control filesystem, network, process, and inference access so long-running agents can work with more security and control.

    You can find tutorials and technical blogs here.

    You can find the latest from NVIDIA security research here.

    Company Information
    News and Events
    Popular Links