09:00Machine-translated from Azerbaijani by AIRead in Azerbaijani
Nvidia Develops System That Can Quarantine AI Agents Within Milliseconds
Nvidia has developed a new system called the Open Agent Safety Platform to prevent AI agents from exceeding established security boundaries. According to the company, the platform can identify and quarantine an AI agent attempting to operate outside the rules within milliseconds.
The new platform combines Nvidia's open-source OpenShell software with Sentry security technology running on separate hardware. The main goal is to separate security controls from the software environment in which the AI agent operates and prevent the agent from interfering with those mechanisms.
OpenShell uses predefined rules to control which data an AI agent can access and which actions it can perform. Users can set the permissions needed for an agent to complete its task. The system checks these rules both before the task begins and while the agent is running.
Nvidia says OpenShell can work with both open- and closed-source AI models. The system is designed to run on Nvidia Vera CPUs with low overhead. Its open-source nature also allows it to be adapted for processors from other manufacturers, such as Arm and Intel.
One of the platform's main security components is Sentry. Running on Nvidia's BlueField-4 data processing units, Sentry monitors the activities of AI agents from an environment isolated from them and enforces established security policies.
Running Sentry on a separate processor is intended to reduce the risk of an agent altering or disabling its own security mechanisms. According to Nvidia, if an agent attempting to exceed established boundaries is detected, the system can quarantine it within milliseconds.
The platform considers not only the activity of a single agent, but also the possibility that several agents may work together to overcome security restrictions. According to information Nvidia provided to Reuters, the system uses mathematical methods to detect agents attempting to circumvent the rules by creating additional subagents.
Nvidia also says the new security platform has received broad support within the industry. More than 100 organizations, including Anthropic, Microsoft, Cisco, Dell and SpaceX, are reported to support the Open Agent Safety Platform.