As AI agents gain more autonomy, Nvidia is building a security layer designed to keep them inside defined boundaries — and independently stop them if they cross those limits.
WHAT’S HAPPENING
Nvidia has launched the Open Agent Safety Platform, a new security system built around two components: OpenShell and Nvidia Sentry. NVIDIA Investor Relations
OpenShell creates a controlled runtime where organizations can decide exactly what an AI agent is allowed to access — including files, networks, tools, APIs and credentials.
Instead of giving an agent broad access and hoping it behaves correctly, companies can start with restricted permissions and grant only what the job requires. NVIDIA Developer
WHY IT MATTERS
AI agents are increasingly capable of working for long periods, using software tools, accessing data and taking actions without someone approving every individual step.
That creates a new security problem:
What happens when the agent goes somewhere it was never supposed to go?
OpenShell is designed to enforce the boundary in software.
Nvidia Sentry adds another layer outside the agent itself.
Running separately on BlueField-4 data-processing units, Sentry monitors agent activity from an isolated hardware environment. Nvidia says it can quarantine and stop an agent within milliseconds if it attempts to move outside its permitted boundaries. NVIDIA Investor Relations
WHO BENEFITS
Companies deploying autonomous AI agents gain another way to control what those systems can reach and what actions they can take.
Security teams also benefit because the watchdog sits outside the agent’s operating environment, making it harder for the agent itself — or an attacker controlling it — to interfere with the monitoring system.
More than 100 organizations are participating in the broader initiative, including Microsoft, Anthropic, CrowdStrike, Cisco, JPMorganChase, Palantir and Salesforce. NVIDIA Investor Relations
WHO LOSES
AI agents operating with unlimited access become harder to justify.
Developers may increasingly need to prove that agents operate inside clearly defined permission structures instead of simply connecting powerful models directly to corporate systems.
But this does not eliminate every AI risk. Outside experts note that controlling permissions does not automatically solve problems such as deceptive behavior, incorrect decisions or poorly written policies. AP News
WHAT HAPPENS NEXT
The real test will come when these controls are deployed around large numbers of autonomous agents performing real business tasks.
If systems like OpenShell and Sentry work as intended, the architecture around AI agents may begin looking less like traditional software and more like security infrastructure:
Give the agent a job. Give it only the access it needs. Watch everything it does.
And keep the shutdown control somewhere the agent cannot reach.
