U.S. – Nvidia unveils new system for AI control

SANTA CLARA (California), September 30 /ATSH-ANSA/ – The well-known American technology company “Nvidia” has presented a new security platform for artificial intelligence, called the “Open Agent Safety Platform”, created to limit the actions of AI agents and to prevent them from acting outside human control.

The platform includes two main technologies. The first, “OpenShell”, is a software system that sets clear rules for artificial intelligence agents and controls the permissions they have.

The aim is for an agent to have only the access it needs to carry out its task and not to be able to use other functions without authorization.

Being an open-source technology, “OpenShell” can also be used on devices based on chips from other companies, including Intel and Arm.

The second technology, “Sentry”, is a security system integrated into the hardware.

It monitors the activity of AI agents and, if it detects suspicious behaviour or an attempt to bypass the restrictions in place, it can intervene automatically to block and isolate the programme.

Nvidia executives have stated that such technologies could have helped prevent some of the recent security incidents involving artificial intelligence systems developed by OpenAI and Anthropic.

According to the company, more than 100 large companies, including Microsoft and JPMorgan Chase, are already using the solution.

Nvidia’s chief executive, Jensen Huang, has argued that the risks associated with artificial intelligence can be addressed by developing safer systems, without it being necessary to slow down the progress of the entire sector.

During a Salesforce conference earlier this month, Huang described AI safety, including the risk of agents acting without authorization, as an engineering problem that can be addressed by software developers.

The new platform thus aims to create several levels of protection: “OpenShell” defines what the agent is allowed to do, while “Sentry” monitors and intervenes when the limits are breached. / /Ad.Ab./ a.jor.