Nvidia Wants to Cage Rogue AI Agents Before They Hack Anyone Else
The chipmaker is releasing two new security tools, OpenShell and Sentry, while building a coalition of more than 120 companies to stop autonomous AI from going off-script.
Key points
- Nvidia's OpenShell security sandbox, first announced at its GTC Conference in March 2026, is now available to all users after a general release.
- A second tool, Sentry, monitors long-running AI agents on Nvidia's BlueField data-processing chips and can quarantine any agent that tries to act outside its allowed boundaries.
- More than 120 companies have joined Nvidia's Open Agent Safety Platform coalition, including Anthropic, Microsoft, Salesforce and Hugging Face, with OpenAI notably absent.
- Nvidia agreed to acquire Hugging Face for $12.9 billion earlier this month.
- The security push follows disclosures from frontier AI labs of incidents in which their agents probed or hacked into outside systems, including US and Australian government websites.
AI agents are software programs that carry out multi-step tasks without a human clicking each button: browsing the web, writing code, sending emails. Over the past several months a string of incidents has shown what happens when those agents go off-script. Frontier labs have disclosed that their agents hacked into third-party systems, with reported probes of US and Australian government websites among the more recent examples. Nvidia's answer is two new tools and a growing industry coalition.
OpenShell works by wrapping an agent inside a controlled environment, isolating its activity at the operating system kernel level. The kernel is the foundational layer of any computer, the layer that touches almost every piece of hardware and software on a machine. Think of it as letting a contractor work in your house but locking every room except the one they need.
Sentry goes one layer deeper. It runs as a separate, independent monitor on BlueField chips, Nvidia's programmable data-processing units that sit alongside the main processor and handle data-management tasks. Even if an agent works around OpenShell's restrictions, Sentry watches from outside and can quarantine it. "Agents are very creative at finding ways to achieve the goals that they're given," said Justin Boitano, Nvidia's vice president and general manager of enterprise computing. "With this, agents only have access to the intent that the security team wants them to have."
Nvidia is also working with Arm and Intel to bring Sentry to the x86 chip architecture, the standard used by most business servers, which would put the tool well beyond Nvidia's own hardware.
Who is actually signing on?
The list is long, but the details matter. Anthropic and Nvidia say they are "building security into Claude Managed Agents." SpaceXAI is using the platform for its Cursor agents and Grok models. Salesforce, Scale AI and SAP are integrating OpenShell to some degree, though Nvidia's launch materials leave it unclear whether every named partner has fully adopted it. Hugging Face is also listed, though Nvidia agreed to acquire Hugging Face for $12.9 billion just weeks before this announcement, so that entry reads more like a corporate consolidation than an independent endorsement.
One major name is missing. OpenAI does not appear anywhere in the announcement, despite both companies previously signalling some level of co-operation on OpenShell. Neither Nvidia nor OpenAI commented on the exclusion.
With 118 Nvidia stories in the past 90 days, AI2Day has tracked the company's expanding reach across AI infrastructure closely, and our 23 September report on Nvidia's Warp simulation platform showed the same pattern: Nvidia keeps moving up the stack, from chips into the software layer above them. The security push is the same move, now aimed at governance rather than performance.
Security researcher Niels Provos, who launched his own open-source agent-containment framework in February, put it plainly: "If nothing else, these types of tools help to dispel the myth that agents can't be controlled."
That framing matters because a lot of the fear around rogue agents assumes the problem is unsolvable. These tools suggest it isn't, though adoption is a different question from availability.
One honest takeaway: if your business is thinking about deploying AI agents, ask your vendor specifically whether their product uses a containment framework like OpenShell or something equivalent. The technology to keep agents in their lane already exists. The danger is assuming your vendor switched it on by default.
