NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment

My take: More than 100 companies, including Anthropic, Microsoft, Salesforce and SAP, joining a platform built to put hard limits on AI agents says a lot about where AI at work is heading. It is no longer just about how well an agent responds, it is about how well you can control it once it starts taking actions on its own.

What NVIDIA is proposing with OpenShell and Sentry is easy to understand: the agent does not have to be trusted to behave, the infrastructure enforces it instead. That matters for any business already handing real tasks (reviewing contracts, moving data, touching internal systems) to an AI agent: security cannot rely on well written instructions alone.

For anyone using AI at work, this confirms something worth keeping in mind: the more autonomy you give an AI tool, the more it matters to know exactly what it can and cannot touch. That is not paranoia, it is good practice.

If you are a business owner or manager about to give an AI agent access to your systems, have you clearly defined the limits of what it can do without supervision?