Nvidia has launched an open security platform aimed at directly restricting AI agents and quickly isolating them if they attempt to overstep their granted permissions, as these systems are increasingly used to perform sensitive tasks semi-autonomously.
The platform is called the NVIDIA Open Agent Safety Platform and consists mainly of the open-source OpenShell software and a reference system called Sentry. OpenShell monitors the agent''s actions and enforces policies during operation, while Sentry uses separate hardware-level monitoring and can quarantine the agent within fractions of a second when it goes outside the specified limits.
Nvidia said that more than 100 entities are involved in developing or using the system, including Anthropic, Microsoft, Salesforce, JPMorgan Chase, Hackingface and SpaceXAI. OpenShell can also be extended to work with computing platforms from other companies, including Arm and Intel.
The company believes that agent security needs controls that extend across software and hardware, rather than relying on restrictions placed within the application alone