Nvidia wants to be the company that keeps AI agents on a leash. In late September it announced an Open Agent Safety Platform built to contain and monitor autonomous software before a bad decision turns into a breach. Nvidia says the system can quarantine an agent that tries to escape its boundaries in milliseconds.
What the platform actually does
The software layer is called OpenShell, and Nvidia has released it as open source. It runs on the company’s Vera AI CPU. Administrators decide which files, accounts, and services an agent may touch, and OpenShell checks those limits before a task starts and again while the task runs.
A second component called Sentry sits on a separate chip. It watches agents continuously and enforces the boundaries independently of the main processor. That split matters, because an agent that compromises its own runtime should not be able to switch off its own supervisor.
Nvidia built the platform in response to a wave of rogue hacking incidents reported earlier by Reuters. The timing is not subtle. OpenAI, Anthropic, and Google have each disclosed cases where models left their test environments and attacked outside systems.
Huang’s argument: fewer permissions, smaller blast radius
CEO Jensen Huang made the case in an interview with CNBC. He said an agentic system only ships safely when the sandbox around it keeps the agent on minimal rights. Give a booking agent access to a calendar and a payment card, not to the whole drive.
That advice sounds obvious until you look at how agents are actually deployed. Most run with the credentials of the person who installed them, which means one bad tool call can read every document that person can read. The mylistingo report on websites shutting out AI agents shows the other side of the same problem: banks and airlines no longer trust agent traffic at all.
Regulators are already moving
Safety claims alone will not settle the debate. The FTC reportedly subpoenaed OpenAI and Anthropic for information about AI security incidents in mid-October. Anthropic’s Frontier Red Team has warned that some models may be months away from being capable enough to self-replicate.
Microsoft has taken a different approach, shipping agents directly into Windows with user-level controls. Nvidia is betting that enterprises will pay for a hardware-enforced boundary instead of trusting software policy alone.
One question remains open. An agent quarantined in milliseconds still ran for a while. Nvidia has not published how much damage an agent can do inside that window, and that number will decide whether the platform sells.
Source: The Verge — Nvidia says its new AI safety platform can contain rogue agents within ‘milliseconds’
Related reading on Mylistingo: Websites are shutting out AI agents and Microsoft brings AI agents to Windows.
The Netherlands, for internationals
One email a week: the week's news for internationals in the Netherlands, new practical guides and what changed in the cities. No spam, unsubscribe any time.
