Nvidia launches a platform to contain AI agents that cross limits
By Automate Basics, written with AI from The Verge's original
Nvidia has launched the Open Agent Safety Platform to restrict what AI agents can access and monitor whether they stay within those limits. It is relevant to teams considering agents that act on work information.
What changed with Nvidia's agent safety platform
Nvidia says its new Open Agent Safety Platform lets users choose what information an AI agent may access. Its OpenShell software checks those restrictions before a task starts and while the agent is working. Nvidia says the platform can quarantine an agent that tries to move beyond its permitted boundaries within milliseconds.
The Verge reports that OpenShell runs on Nvidia's Vera AI CPU. The platform also uses Nvidia's Sentry technology on a separate chip to keep watch over agents and enforce their limits. For a workplace team, the distinction is between telling an agent what it may use and having a system that checks whether it follows those limits during the task.
Who can use the Open Agent Safety Platform?
The Open Agent Safety Platform is relevant to organizations giving AI agents access to information for work tasks. Nvidia's approach is to limit that access to what each agent needs, then monitor whether the agent stays within those limits. The Verge reports that Anthropic, Microsoft and SpaceX are among the companies backing the platform.
The source does not say who can obtain the platform, which plans include it or what it costs. It also does not describe a sign-up process or name workplace products that already include these controls. Teams interested in using it should treat availability and purchasing details as open questions, rather than assume it is part of an AI tool they already use.
How do I assess it for an agent at work?
A useful first task to consider is one where an agent needs a limited set of work information, not unrestricted access. Before assigning the task, identify what information the agent genuinely needs and what should remain outside its reach. That gives a team a concrete boundary to ask about when evaluating Nvidia's platform.
Nvidia says OpenShell checks access restrictions before and during a task, while Sentry continues to monitor the agent. Ask how a proposed setup would apply those controls to the task, and check whether the agent's output relies only on information it was permitted to use. The Verge's report does not provide instructions for setting up or testing the platform, so those checks are evaluation questions, not a claim that its controls have been independently verified.
Frequently asked questions
What does Nvidia's Open Agent Safety Platform do?
Nvidia says users can set limits on the information an AI agent may access. Its OpenShell software checks those limits before and during a task, while Sentry monitors the agent on a separate chip. Nvidia says the platform can quarantine an agent that tries to cross its boundaries within milliseconds.
Is the Open Agent Safety Platform available to my team?
The Verge reports Nvidia's launch but does not say who can obtain the platform, whether it is included in any plan or what it costs. Teams would need those details before deciding whether they can use it with their work agents.
Why would an organization restrict an AI agent's access?
An agent may need some work information to complete a task without needing access to everything else. Nvidia's platform is designed to enforce a chosen boundary and monitor whether the agent stays inside it. A team can start by identifying the information required for a specific task.
Related reading
Thoughtful Things launches AI sound sampler Engram on KickstarterThoughtful Things has launched a Kickstarter campaign for Engram, a sampler that uses AI to alter recorded audio and generate unusual sounds. It is aimed at musicians and sound designers experimenting with audio, not people seeking finished songs.
Google tests Flipkart checkout inside Gemini and AI Mode in IndiaGoogle is testing purchases from select Flipkart listings inside Gemini and AI Mode in India, TechCrunch reports. The limited test may interest teams that use AI to research work purchases, but most users cannot try it yet.
Meta outlines Mac computer use and connectors for Muse AI appMeta announced plans to add Mac computer use, connectors and a dedicated email address to its Muse AI app, according to TechCrunch. The features could make the personal agent more useful for work, but Meta has not said when they will arrive.
Written with AI from The Verge's original and published after automatic checks: every figure here appears in the original, and no sentence is copied from it. The picture is AI-generated. The original is the authority.
Source: Nvidia says its new AI safety platform can contain rogue agents within ‘milliseconds’, The Verge, 28 Sept 2026.