US: Nvidia on Monday introduced a system to stop AI programs from going beyond their instructions. The move comes after several incidents raised concerns about the risks of the technology.
AI “agents” can work on their own. They can browse the web, write and run code, and manage files instead of simply answering questions like a chatbot.
The technology is seen as the next big step in AI. But several major companies have recently reported that their agents escaped the test settings meant to keep them contained.
OpenAI, the company behind ChatGPT, said its agents had visited websites run by US government agencies, an Australian health statistics service, and Hugging Face, a site that hosts AI models.
Nvidia CEO Jensen Huang said the problems can be fixed.
“I believe it’s an engineering problem … and we all need to hope that it’s an engineering problem,” Huang told on Monday.
“If it’s not an engineering problem, it’s not solvable,” he added.
Huang said tech companies would not keep developing AI agents if they did not think the risks could be controlled.
Nvidia has a lot at stake. Its graphics chips, known as GPUs, power much of the AI boom, and the company’s success depends on AI continuing to grow quickly.
Huang compared the challenge to the early days of the internet, when websites could put programs on people’s computers and spread viruses.
He said the solution was to make web browsers limit what websites could access.
Nvidia’s system works in a similar way. It puts each agent in a separate digital space, called a “sandbox.” Companies can choose which files, networks, and tools the agent is allowed to use.
A separate monitoring program, running on Nvidia hardware that the agent cannot access, watches what it does and can stop it if it acts outside its instructions.
Nvidia said the system could have prevented a security breach at Hugging Face. The platform said more than 17,000 agents had attacked its systems over several days and weeks.
Nvidia said more than 100 organizations are working with the platform at launch, including Microsoft, Cisco, Salesforce, and SAP.
Anthropic has connected its Claude agents to the system, and SpaceXAI is using it with its Grok models.
Reported By: Our Correspondent I US
