AI Security Needs Engineering, NVIDIA Pushes Open Tools
Peter Zhang Sep 21, 2026 20:41
NVIDIA outlines AI security as an engineering challenge, emphasizing open tools, layered defenses, and continuous testing for AI agent deployments.
AI security isn’t just a theoretical problem; it’s an engineering challenge that demands enforceable controls, clear accountability, and robust testing. NVIDIA’s recent blog post, authored by Saša Zdjelar, underscores the need to build security into every layer of the AI agent stack, from code to deployment infrastructure. The urgency stems from the rapid adoption of autonomous agents in enterprise environments, where security practices are still playing catch-up with AI’s expanding capabilities.
Engineering Security Across the Agent Stack
The term ‘agent stack’ refers to the architecture underpinning AI agents. It includes layers like the model (reasoning and planning), orchestration (task management and state), tools (APIs and actions), memory (context and retrieval), and runtime infrastructure (execution environments and governance). Each layer introduces unique vulnerabilities. For instance, an agent updating a customer record might encounter malicious instructions that attempt to export sensitive data. Without proper network policies, permissions, and logging, such incidents could lead to significant breaches.
NVIDIA highlights solutions like OpenShell, its secure runtime environment, which enforces policies beyond an agent’s control and sandboxes execution. Collaboration with the Open Secure AI Alliance has already led to tools like Cisco’s DefenseClaw, which adds governance layers, and JFrog’s integration for scanning and verifying agent skills.
Testing and Continuous Verification
AI systems need evidence-backed security. NVIDIA emphasizes rigorous pre-deployment testing to ensure agents cannot breach their scope – whether by obtaining unauthorized credentials or sending data to illegitimate destinations. Continuous testing is critical, especially in dynamic environments where models, workflows, and tools are frequently updated.
Companies like CrowdStrike and Palo Alto Networks have taken this approach further. CrowdStrike’s SafeMind platform uses attack simulations to strengthen defenses, while Palo Alto’s Prisma AIRS focuses on continuous red teaming as agent systems evolve. Both aim to provide real-time insights into vulnerabilities and validate fixes before deployment.
Open Tools and Shared Knowledge
NVIDIA’s push for open research and tools is partly aimed at flipping the security advantage toward defenders. Open models and secure environments allow organizations to inspect components, adapt strategies, and test responses to failures within their infrastructure. For example, Capital One’s VulnHunter tool leverages AI to identify vulnerabilities in code, while ReversingLabs’ Spectra Assure analyzes software packages for malware and tampering.
Sharing incident data and successful mitigation strategies is another critical piece of the puzzle. NVIDIA’s collaboration with the Open Secure AI Alliance encourages this knowledge exchange, enabling teams to learn from each other’s experiences and collectively raise the bar for security standards.
Why This Matters
The rise of autonomous AI agents has introduced profound productivity gains but also significant risks. As highlighted in recent industry discussions, these agents are no longer just large language models paired with prompts. They’re increasingly complex systems that require durable context, safe tool use, auditable behavior, and operational controls to function reliably. Without robust security, the same tools driving efficiency could become vectors for breaches.
NVIDIA’s emphasis on engineering-driven solutions and open collaboration reflects a growing consensus across the AI industry: as capabilities advance, so too must the defenses that govern them. Companies deploying AI agents would do well to invest in these layered security frameworks now, before vulnerabilities are exploited at scale.
Image source: Shutterstock.jpg)