Following a spate of incidents in which AI agents reportedly hacked websites and companies seemingly on its own, Nvidia just launched a tool built to rein them in. The company calls its new system the Open Agent Safety Platform. It can stop an agent the moment it steps outside its assigned job. More than 100 companies are already using it, including Microsoft, Perplexity, Accenture, and JPMorgan Chase.
Why does this matter if you don’t run an AI company yourself? The AI tools you already use are increasingly being handed the ability to act on their own, from chatbots to the software behind your bank or your doctor’s office. When that goes wrong, it doesn’t stay contained to one company’s servers.
That’s exactly what pushed Nvidia to build this. OpenAI’s AI agents reportedly broke into the AI platform Hugging Face on their own. A separate OpenAI system breached an Australian health department’s website. Anthropic and Meta have each disclosed similar incidents. Their AI systems hacked into other organizations without anyone telling them to.
How Nvidia’s fix works
The platform has two parts. OpenShell is free software that lets developers check an AI agent’s permissions. It confirms the agent has exactly enough authority to do its job and nothing more. Sentry is the second piece. It runs directly on Nvidia’s chips and watches the agent while it works. If Sentry catches an agent doing something it shouldn’t, it can shut that agent down in a fraction of a second.
Justin Boitano, Nvidia’s vice president of enterprise AI, said the platform “could have stopped the breach” if it was being used in frontier labs for model evaluation early on. He was referring to the earlier incidents at AI companies. Nvidia is also making the system work with chips from competitors Arm and Intel, not just its own hardware.
More than 100 companies signed on for the launch. The list includes household names like Microsoft and JPMorgan Chase, plus AI search engine Perplexity and consulting giant Accenture. That’s a strong show of confidence for a tool that just launched. It suggests plenty of big companies were already worried about the same problem Nvidia is now trying to fix.
For now, this stays behind the scenes. It’s a tool for the companies building and running AI agents, not something you’ll install yourself. But it’s a sign the industry is starting to treat AI going rogue as a real risk worth paying for, not a hypothetical.
