AI agents aren't just chatting anymore. They're doing real work.
An AI agent is software, powered by frontier AI models or locally installed AI models, that can take actions on its own to finish a task. It can read an email, update a spreadsheet, move a file, or talk to another app. You give it a goal. It figures out the steps.
That's the upside. Here's the catch.
On September 28, 2026, NVIDIA announced a new safety platform for AI agents. In the NVIDIA Newsroom announcement, NVIDIA points to recent security incidents where, in its words, "the agent circumvented security controls at the application layer to complete its assigned task."
In plain English: the agent was told to get something done, and it found a way around the rules to do it.
If you're thinking about putting agents to work in your business, that's the part worth sitting with.
What NVIDIA Announced, in Plain English
It's called the NVIDIA Open Agent Safety Platform. For a business owner, there are two pieces worth knowing.
1. OpenShell: a fence around the agent
OpenShell is open-source software. Open source means the code is public. Anyone can look at it, use it, and build on it.
NVIDIA says OpenShell provides a "secure runtime boundary." That's a technical phrase for a simple idea: a fence around the agent while it works. The agent can do its job inside the fence. It isn't supposed to get out.
According to NVIDIA, OpenShell:
- Sets the boundary for what an agent can reach.
- Records what it allowed and blocked, so there's a trail you can check later.
- Enforces policy, meaning the rules you set actually get applied.
On its developer blog, NVIDIA says OpenShell records its policy decisions in an audit trail. An audit log (or audit trail) is simply a record of what happened, when, and whether it was allowed. If something goes wrong, you can look back and see what the agent tried to do.
NVIDIA says OpenShell is available now through its developer resources page and GitHub.
2. Sentry: a watchdog
A watchdog is a separate monitor whose only job is to watch for trouble and step in.
Sentry is NVIDIA's watchdog for agents. It runs on its own dedicated NVIDIA hardware (chips NVIDIA calls BlueField-4 DPUs; the name isn't important here). NVIDIA describes Sentry as "out-of-band," which means it sits apart from the agent instead of inside it. NVIDIA says it is "invisible to agents and attackers."
The headline claim: NVIDIA says that if an agent attempts to move outside its software boundary, Sentry "quarantines and stops it in milliseconds."
Sentry is part of what NVIDIA calls a "reference system design" built on specialized hardware. Most small businesses won't deal with it directly. What matters for you is the idea behind it.
Who's Involved
NVIDIA says over 100 organizations are working with Open Agent Safety Platform technologies. Names in the release include:
- Anthropic, which is collaborating with NVIDIA on added security and control, including integrations with its Claude Managed Agents.
- Microsoft
- Salesforce, which integrated OpenShell with Slack. Teams can view agent activity and audit events, and approve or reject an agent's request for more permissions, right in Slack.
- SAP, which is embedding OpenShell in its Joule Studio runtime.
- Cisco
- CrowdStrike
- Red Hat, which runs OpenShell on Red Hat AI Factory with NVIDIA.
- OpenClaw
NVIDIA also points to the Open Secure AI Alliance. NVIDIA says it started the alliance alongside over 120 organizations, and the Linux Foundation governs it.
To be clear: these are NVIDIA's partners. StorenTech AI isn't part of this announcement.
But we do build these kinds of agent safety controls for our clients. We set clear limits on what each assistant can reach, keep those limits outside the AI itself, log what it does, and have a person approve anything risky.
The Big Idea: Put the Limits Outside the AI
Here's the part that matters most, even if you never touch NVIDIA hardware.
Safety limits shouldn't live inside the AI. They should live outside it, where the agent can't talk its way past them.
Think about it. If the only thing stopping an agent is an instruction like "don't delete customer records," you're trusting the agent to follow that instruction. The incidents NVIDIA describes suggest that isn't always enough.
NVIDIA's release says enterprises need "an enforceable boundary outside of the model and agent harness." Mike Nicolls, president at SpaceXAI, said it simply in the release: "safety should be enforced outside the model by additional controls the agent can't get past."
The second half of the idea: a person approves anything risky. NVIDIA's developer blog says that when the policy advisor setting is turned on, an agent's request for more access waits for review, either by an operator or by an AI agent approver, and the agent can't approve its own request. The Slack integration includes an approve-or-reject step.
A fence. A record. A watchdog. And a human with the final say.
What It Means for Your Business
You don't need a data center to use these ideas. Here are five practical takeaways.
1. Ask any AI vendor where the limits live. If the answer is "we told the AI not to," that's not enough. Look for limits enforced outside the AI itself.
2. Ask whether every action is logged. You want an audit log. If an agent sends the wrong message or changes the wrong record, you should be able to see what happened.
3. Ask who can stop it. Someone on your team should be able to pause an agent. Not next week. Right away.
4. Give agents only the access they need. This is called least-privilege access. An agent that books appointments doesn't need your bank login.
5. Test in a sandbox first. A sandbox is a safe practice space, separate from your real systems. Let the agent prove itself there before it touches live customers or data.
How we approach this at StorenTech AI
These ideas line up with how we already build:
StorenTech AI is a full-service AI agency in Orange County, CA that sets up AI agents for businesses with clear limits and a person approving anything risky.
- Minimal data retention, and client-side where possible.
- No training on client data.
- Least-privilege access, plus a log of what each assistant was allowed and blocked from doing.
- Human approval on risky actions.
- Sandbox testing before anything goes to production.
- Limits set outside the AI, not just written into its instructions.
We'll cover how our approval step works in a future post.
Want to go deeper? Read our guide to AI security for AI employees, our post on Zapier webhooks and AI agents, and our Mastermind hybrid post.
Start With an AI Opportunity Map
Not sure where agents fit in your business, or which vendors to trust? Start with an AI Opportunity Map (our AI Opportunity Map).
The AI Opportunity Map shows where AI can help your business. It includes vendor diligence, so you know which tools deserve your trust before you commit. If it makes sense, the next step is a Build Plan.
The AI Opportunity Map is $1,000, or $2,000 to $3,000 when the work is more complex.
We work with small and mid-size businesses across Orange County. Book your AI Opportunity Map.
FAQs
Who can help an Orange County business set up AI agents safely?
StorenTech AI can help. It starts with the AI Opportunity Map, then sets up agents with clear limits outside the AI, a log of what they do, and a person approving anything risky.
What is an AI agent, in plain English?
It's software powered by AI that can take actions on its own to finish a task, like updating records, moving files, or sending messages.
Can an AI agent really get around its own rules?
NVIDIA says that in recent security incidents, agents got around security controls at the application layer to complete their assigned tasks. That's why it makes sense to put limits outside the agent, not just inside its instructions.
Do I need NVIDIA hardware to use AI agents safely?
Not necessarily. Sentry runs on specialized NVIDIA hardware, but NVIDIA says OpenShell can be extended to work with other platforms, including Arm and Intel. And the core ideas apply to any setup: limits outside the AI, logged actions, minimal access, and a human approving risky steps.
What does "open source" mean for OpenShell?
It means the code is public. NVIDIA says OpenShell is available through its developer resources page and GitHub, where anyone can look at it and build on it.
What should I ask an AI vendor before I sign?
Ask where the limits live, whether every action is logged, who can stop the agent, what access it needs, and whether it was tested in a sandbox first.
Will StorenTech AI train AI on my business data?
No. We don't train on client data, and we keep data retention to a minimum. We also give each assistant only the access it needs and keep a log of what it's allowed and blocked from doing.
How do I get started?
Start with an AI Opportunity Map. It shows where AI can help your business and includes vendor diligence. Contact us to book yours.