NVIDIA announced its Open Agent Safety Platform on 28 September, combining software controls with a hardware-based monitoring design to restrict what autonomous AI agents can access and do. The launch addresses a practical problem for businesses: an agent trusted to complete a task may also need limits that it cannot change for itself.
“Safety and security require full-stack engineering,” NVIDIA founder and chief executive Jensen Huang said in the company’s announcement. Its approach places enforcement around the systems running an agent, extending oversight beyond instructions given to the AI model. That distinction matters when software can act on business systems rather than simply suggest what a person should do.
The software component, OpenShell, creates a controlled workspace for an agent and applies permissions outside that workload. NVIDIA’s technical documentation describes controls over files, processes and outbound requests, alongside credential protection. Its supervisor checks requests against policy, while credentials can remain outside the agent’s own working environment.
For example, a business could design a support agent that reads approved product information and prepares a response, while withholding permission to change customer accounts. This is an illustration of the permission model, not a verified customer deployment. The operational benefit would depend on configuring the permitted actions correctly and ensuring that every relevant route to a system is controlled.
The additional hardware layer, Sentry, is designed to monitor agents from a separate part of the infrastructure. NVIDIA describes an optional watchdog running on BlueField-4 data processing units, using its DOCA software to connect agent activity with identity and access policies. The company says this separation gives operators an enforcement point beyond the agent’s reach.
NVIDIA says agents crossing their boundaries can be quarantined in milliseconds, a claim that Techord has not independently tested. OpenShell software is available, but that should not be read as confirmation that every hardware configuration or partner integration is ready for an Australian buyer. The announcement does not establish a local implementation price or delivery timetable.
Australian organisations already have reasons to examine those surrounding controls. In a 28 September ASD advisory, the Australian Signals Directorate reported unauthorised access to organisational AI services through compromised credentials, sessions and other weaknesses. It recommended accountable ownership, minimum necessary access and stronger protection of accounts and API keys.
The agency’s separate harness guidance addresses the software around a model that manages tools, state and actions. Its recommendations include human oversight of high-impact actions, continuous monitoring and phased deployment. These are broader operational practices, not an endorsement of NVIDIA’s product or proof that installing it satisfies a business’s obligations.
The usefulness of restrictions also needs testing against real work. Somesh Jha, a University of Wisconsin computer science professor, told AP that balancing security with useful agent behaviour needed evidence: “This can only be answered using case studies.”
For businesses evaluating the platform, the immediate decision is which actions an agent should be allowed to perform and which require human approval. NVIDIA is offering technology to enforce that boundary, but organisations still have to define it, test it and decide how they will respond when an agent reaches it.
