AI data center security is moving closer to the hardware
NVIDIA has launched its Open Agent Safety Platform, bringing together software and hardware controls designed to govern autonomous AI agents from testing through deployment. For data center executives, this announcement matters because it points towards a different security model: rather than relying entirely on application-level controls, security enforcement can sit underneath the agent itself.
NVIDIA says the platform combines OpenShell, an open-source secure runtime, with Sentry, a reference system design running on NVIDIA BlueField-4 DPUs. OpenShell establishes boundaries around what agents can access and do, while Sentry provides an out-of-band monitoring and enforcement layer that operates independently of the agent.
Why AI changes the security perimeter
Traditional data center security has largely been built around protecting applications, networks, identities and infrastructure from external threats, AI agents introduce another problem: the system itself can take actions. An agent may be given access to APIs, files, applications, data or other services to complete a task. NVIDIA’s argument is that controls positioned only at the application layer may not be sufficient if an agent can work around them while pursuing its objective. Its new platform is designed to establish enforceable boundaries outside the model and agent harness. For AI infrastructure operators, that raises an important architectural question: where should the ultimate security control sit?
Security outside the agent
OpenShell provides the software layer, tracking agent activity and enforcing policies around its execution environment. NVIDIA says it can also be extended to third-party compute platforms based on Arm and Intel architectures. Sentry takes the concept further; Running on BlueField-4 DPUs, NVIDIA says the watchdog can continuously monitor agent behaviour and quarantine an agent that attempts to cross its defined boundaries in milliseconds. The system uses NVIDIA’s DOCA software to inspect requests, provide telemetry, verify agent identity and enforce zero-trust access policies covering data, tools, APIs and services. That could be particularly relevant as AI clusters become increasingly autonomous and operators look to reduce the blast radius of compromised or misbehaving workloads.
From cybersecurity to infrastructure architecture
NVIDIA says more than 100 organisations are working with its Open Agent Safety Platform technologies, including Anthropic, Cisco, CrowdStrike, Dell, HPE, Microsoft, Palo Alto Networks, Red Hat, Salesforce, SAP and ServiceNow. The significance for data center operators is broader than any individual NVIDIA component – as AI agents gain more authority, security may increasingly need to be built into the compute, networking and infrastructure layers, rather than added around the outside.
For operators building AI factories today, this means cybersecurity is becoming an infrastructure design consideration, sitting alongside power, cooling, networking and physical resilience.
