Security and performance move in lockstep, not against each other
As enterprises race to embed artificial intelligence into the core of their operations, a quieter but equally consequential question has emerged: who is watching the machines that watch everything else? On Tuesday, cloud security firm Upwind and chip giant NVIDIA announced a deepened partnership aimed at answering that question — building protections into GPU-powered AI infrastructure that operate in real time, without slowing the systems they guard. The collaboration reflects a broader reckoning in the technology world, where the speed of AI adoption has outpaced the maturity of AI-specific security, leaving organizations exposed to threats that traditional tools were never designed to see.
- Enterprises deploying AI at scale face a new class of cyber threat — prompt injections, jailbreaks, and model theft — that conventional security tools are fundamentally blind to.
- The urgency is compounded by speed: organizations are racing to adopt large language models and autonomous agents faster than security frameworks can evolve to protect them.
- Upwind and NVIDIA are responding by embedding runtime visibility directly into GPU infrastructure, including NVIDIA's DGX systems and Blackwell architecture, so threats are caught as they happen rather than discovered after the damage is done.
- The integration of NVIDIA's Garak framework gives security teams a way to continuously stress-test AI systems against real-world attack scenarios before adversaries can exploit them.
- The partnership is landing as a proposed industry standard — trusted AI — where security and performance are designed to advance together, not sacrifice one for the other.
Upwind, a Tel Aviv-based cloud security company, announced Tuesday that it is expanding its collaboration with NVIDIA to protect AI workloads running on GPU-powered infrastructure. The driving concern is one that has grown louder as enterprises scale their AI deployments: how to defend these systems against a new generation of cyber threats without degrading the performance that makes them valuable.
The partnership operates on two tracks. Upwind is using NVIDIA's NIM microservices to sharpen its own internal security capabilities — improving how it analyzes runtime data, correlates vulnerabilities, and models threats across complex systems. Simultaneously, Upwind is building dedicated protections for NVIDIA's hardware platforms so that AI workloads benefit from continuous runtime visibility, severity-ranked risk signals, and zero measurable performance impact.
The threats in question are specific to AI and largely invisible to traditional security tools. Large language models and autonomous agents introduce attack surfaces that didn't exist before — prompt injection attacks that manipulate model behavior, jailbreaks that bypass safety guardrails, and attempts to extract training data or model weights. Upwind's approach is to monitor what happens inside a system as it runs, catching anomalies in real time rather than reconstructing events after a breach.
The collaboration also incorporates NVIDIA Garak, an open-source framework that simulates AI-specific attacks, allowing enterprises to continuously validate whether their systems can withstand real-world pressure. Combined with Upwind's monitoring of API calls and runtime behavior, organizations gain both a testing layer and a live defense layer working in concert.
Leadership from both companies framed the partnership as foundational rather than incremental. NVIDIA's vice president of cybersecurity stressed that security must be designed in from the start, while Upwind's senior vice president of platforms described the goal as enabling organizations to deploy AI at scale with full confidence. The ambition is a new benchmark for the industry — one where trusted AI means security and performance are no longer in tension, but advancing together.
Upwind, a cloud security company based in Tel Aviv, announced on Tuesday that it is deepening its work with NVIDIA to protect artificial intelligence workloads running on GPU-powered infrastructure. The partnership addresses a problem that has become urgent as enterprises rush to deploy AI systems at scale: how to keep those systems safe from cyber threats without slowing them down.
The collaboration works in two directions. Upwind is using NVIDIA's NIM microservices—modular AI components—to power its own internal security operations, improving how it analyzes runtime data, correlates vulnerabilities, and models threats across large systems. At the same time, Upwind is building dedicated protections for NVIDIA's hardware platforms, including the DGX system and the newer Blackwell architecture, so that AI workloads can run with continuous visibility into what's happening at runtime, with risks ranked by severity, and without any measurable impact on performance.
The security challenge is real and specific to AI. As organizations adopt large language models and autonomous agents, they create new surfaces for attack. Threats include prompt injection attacks—where malicious input tricks an AI model into behaving badly—jailbreaks that circumvent safety guardrails, and attempts to steal training data or model weights. Traditional security tools were not built to detect these kinds of attacks. Upwind's approach is to watch what happens inside the system as it runs, rather than waiting for a breach to be discovered after the fact.
The partnership introduces what both companies describe as five key advantages. The first is performance: by using NVIDIA's accelerated computing, Upwind can process security data faster. The second is flexibility in where systems can be deployed—whether in public clouds, private data centers, or sovereign cloud environments where data must stay within a country's borders. Third is cost: the ability to scale AI inference and analytics without proportional increases in spending. Fourth is privacy: strict enforcement of where data lives and who can access it. Fifth is customization: engineering tailored to how a particular customer wants to run AI.
Upwind has also integrated NVIDIA Garak, an open-source testing framework, into its security validation layer for large language models. Garak simulates the kinds of attacks that worry security teams—prompt injections, jailbreaks, data theft—so that enterprises can continuously test whether their AI systems will hold up under real-world pressure. By combining Garak's testing with Upwind's ability to watch API calls and runtime behavior, organizations can validate that their AI applications remain safe and compliant even as they evolve.
Dan Yahav, senior vice president of platforms at Upwind, framed the partnership as a way to let organizations "deploy AI at scale—safely, efficiently, and with full confidence." Ariel Levanon, vice president of cybersecurity at NVIDIA, emphasized that "developers must design systems that are secure from the start," and positioned this collaboration as a step toward that goal. The partnership sits within Upwind's broader strategy around AI security, which includes runtime protection for AI workloads, vulnerability management specific to AI systems, and API security aware of how language models behave. Together, the two companies are trying to establish what they call a new standard: trusted AI, where security and performance advance together rather than trade off against each other.
Citações Notáveis
By combining Upwind's runtime visibility and protection with NVIDIA's accelerated AI infrastructure, we're helping organizations deploy AI at scale—safely, efficiently, and with full confidence.— Dan Yahav, SVP Platforms at Upwind
Developers must design systems that are secure from the start. By incorporating NVIDIA's accelerated computing and security-ready infrastructure, Upwind is changing how organizations understand and defend modern cloud environments.— Ariel Levanon, VP of Cybersecurity at NVIDIA
A Conversa do Hearth Outra perspectiva sobre a história
Why does AI infrastructure need its own security approach? Isn't traditional cloud security enough?
Traditional security watches the perimeter and the network. But AI systems create new attack surfaces—you can manipulate a language model through the prompts you feed it, or trick it into revealing its training data. These threats don't show up in network logs the way a port scan does. You need to watch what the model is actually doing, in real time.
So Upwind is essentially adding a security layer that understands AI-specific attacks.
Exactly. And they're doing it without slowing things down. That's the hard part. When you add monitoring to a system, you usually pay a performance cost. Upwind claims zero impact, which matters when you're running inference at scale—every millisecond counts.
What does it mean that they're using NVIDIA's own tools—NIM and Garak—to do this?
It's a smart move. NIM microservices are NVIDIA's building blocks for AI applications, so Upwind is using the same language, so to speak. And Garak is NVIDIA's own testing framework for finding weaknesses in language models. By using it, Upwind gets credibility and alignment with what NVIDIA customers already understand.
The announcement mentions five key advantages. Are those real technical wins or marketing language?
They're real, but they're also the things every enterprise wants to hear. Performance, flexibility, cost, privacy, customization. The substance is that this partnership lets you run AI safely in different environments—public cloud, private data center, sovereign cloud—without compromising on security or speed.
Who actually benefits most from this?
Large enterprises deploying AI at scale. If you're running one small language model, you probably don't need this level of sophistication. But if you're NVIDIA's customer—if you've bought DGX systems or Blackwell chips—and you're running multiple AI workloads in production, you need to know they're protected and compliant. That's where Upwind comes in.