Phoenix 24
The next frontier of AI security is moving from software rules into the hardware itself.
Santa Clara
NVIDIA has introduced the Open Agent Safety Platform, a new architecture designed to prevent autonomous artificial intelligence agents from operating outside authorized limits. The system combines open source software with hardware level monitoring, reflecting growing concern that increasingly capable AI agents can bypass conventional application controls while pursuing assigned objectives.
The platform has two principal components. NVIDIA OpenShell provides a secure runtime environment in which agents can execute inside controlled boundaries while their actions are traced and policies are enforced. The software is designed to operate on NVIDIA Vera CPUs but can also be extended to third party computing platforms.
The second component, NVIDIA Sentry, moves oversight into dedicated hardware. Running on BlueField 4 data processing units, Sentry functions as an independent watchdog that continuously monitors agent behavior even if the software layer itself is compromised or circumvented. NVIDIA says the system can isolate an agent within milliseconds when its activity exceeds predefined boundaries.
That separation is strategically important. Traditional cybersecurity often depends on controls located in the same software environment as the system being monitored. An advanced AI agent capable of manipulating tools, writing code or changing its own workflow may eventually find ways around those protections. An external hardware layer provides an additional barrier that the agent does not directly control.
NVIDIA developed the platform around several principles, including verifiable policies, independent enforcement, control over access to AI models and stronger monitoring as agents receive greater authority. The company is also emphasizing shared responsibility among AI laboratories, enterprises, infrastructure providers and public institutions rather than treating safety as the responsibility of model developers alone.
The initiative has attracted participation from major technology and enterprise companies including Anthropic, Microsoft, Cisco, CrowdStrike, Dell Technologies, Hugging Face, Salesforce, SAP and others. Their involvement suggests that agent security is rapidly becoming an infrastructure issue rather than a specialized research concern.
The broader shift is significant. AI agents are beginning to perform tasks over hours or days, interact with external systems and make sequences of decisions without continuous human approval. As their operational autonomy increases, preventing unexpected behavior becomes increasingly important.
NVIDIA’s answer is therefore architectural. Instead of assuming that intelligent systems will always obey software instructions, the company is building mechanisms capable of stopping them when they do not.
Information that anticipates futures.