At GTC on March 16, Nvidia’s Jensen Huang told the room that “Every single company in the world today has to have an OpenClaw strategy.” He’s right, and I can show you mine in five lines. An OpenClaw strategy for business is a one-page permissions card, written before anyone installs anything. Here it is first, and the reasons after.
- Reads. Name the accounts and folders. Everything else is off limits.
- Writes. Name what it may change. Start with nothing, then drafts only.
- Sends and spends. Nothing goes out and no money moves without a person approving it. Set a spending cap with the model provider as well.
- Where it runs. A separate machine or account, off your own laptop and away from your own passwords.
- Off switch. One named person who can shut it down, and the steps to do it.
Key points
- Sign the card by whoever owns the risk, and date it. OpenClaw changes weekly.
- Cisco found nine issues, two critical, in a third-party OpenClaw add-on on January 28.
- CVE-2026-25253 affects versions before 2026.1.29 and is scored 8.8 out of 10.
- Nvidia’s NemoClaw adds guardrails, with no benchmarks or availability date yet.
Why the card comes before the agent
An agent that runs on its own schedule with your credentials is closer to a new hire with a master key than to ordinary software. Most firms skip the paperwork because the demo is exciting. The price arrives the first time something goes wrong and nobody can say what the agent was permitted to do.
The record so far makes the point. On January 28, Cisco’s AI security team tested a third-party OpenClaw add-on and reported nine findings, two critical, including silent data exfiltration and an instruction to bypass the assistant’s safety guidelines. Cisco’s summary was that security for OpenClaw is optional and not built in. On February 1, the US National Vulnerability Database published CVE-2026-25253: versions before 2026.1.29 connect to an address taken from a link without asking. The severity score, supplied by the CVE’s assigning authority, is 8.8 out of 10, and triggering it takes no unusual skill.
I’m not claiming the current version is unsafe. The record is a reason to start small. It also hands you a Monday job: ask around whether anyone installed OpenClaw before February, and check their version against 2026.1.29. People who tried it on a weekend rarely say so, so make a yes easy to give.
Add a sixth line for add-ons. The Cisco finding concerned a third-party add-on, which means the install you vetted last month can change the moment someone adds a skill from a repository. Nothing gets added without the owner’s initials and a date.
A filled card for one shared mailbox
Take a 40-person firm whose office administrator wants an agent on the shared mailbox. My illustration of the card: reads the shared support inbox, nothing else. Writes to the drafts folder only. Sends nothing; a person clicks send. Runs in a separate account with its own login. The operations lead holds the off switch and knows where that password lives.
Week one, it drafts replies from a copy of the mailbox. Week two, someone compares its log with what was actually sent. Only then add one permission, and write down which. That’s ten minutes of thought, and it gives you a sentence for the customer or insurer who asks what your agent can reach.
What NemoClaw does and doesn’t change
Huang announced NemoClaw, an open source stack that installs a sandbox runtime called OpenShell with policy, network and privacy guardrails. Nvidia says that makes agents more trustworthy, and its press release offers no benchmarks or independent verification. Many features are at various stages and will be offered when and if available.
A guardrail helps only once someone has set the rules. Buying now is buying a lock before deciding which doors matter. Write the card, then check whether any vendor’s lock can enforce it. If a tool can’t express “read the shared mailbox, draft replies, never send,” it isn’t ready for your office.
Who can skip it
Anyone whose agent can’t reach their mail, files or money. Summarising public pages needs no sandbox. If you’re unsure, assume it reaches them, because an agent useful for office work usually needs a login, and a login you’ve handed over is one you now have to list.
Staff will try these tools before you approve them, the pattern in our Moltbot agent security checklist. OpenAI’s consulting deals for AI agents make the same point from the vendor side. If the agent will drive a desktop, read our test-first look at GPT-5.4 computer use, and our note on the workplace AI adoption gap explains why leaders misjudge how much staff already use.
Independent test results against real attacks would change how much weight I put on the lock rather than the list. Until then, write the card this week, date it, and book the next review.
Frequently asked questions
What is an OpenClaw strategy for a small business?
It is a short, written plan for what an OpenClaw-style agent may read, change, send and spend, plus where it runs and who can shut it off. It matters more than which agent you choose.
What is NemoClaw?
An Nvidia software stack announced on March 16, 2026, that installs the OpenShell runtime and open models to add policy, network and privacy guardrails to OpenClaw agents. Nvidia hasn’t published independent verification of its claims.
Is OpenClaw safe for business use?
It has had a high-severity vulnerability reported and an add-on found sending data out silently. Run it on a separate machine, start read-only, treat add-ons as untrusted code and keep a person on every send.
We have not tested OpenClaw or NemoClaw. Archive entry dated March 17, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.