11:17Machine-translated from Azerbaijani by AIRead in Azerbaijani
Nvidia Unveils Platform That Quarantines AI Agents in Milliseconds
On September 28, Nvidia unveiled a security platform called the Open Agent Safety Platform for keeping AI agents under control. AI agents are systems that perform assigned tasks without human intervention. The platform is designed to prevent them from exceeding preset boundaries. According to Nvidia, the system can quarantine agents that break the rules within milliseconds.
One of the platform's main components is the open-source OpenShell software. Users choose in advance which data and systems an AI agent will be able to access. OpenShell checks these restrictions before a task begins and while it is being performed. The software runs on Nvidia's Vera AI processor and creates a secure test environment for the agent that is isolated from external systems. Another protection system, called Sentry, monitors the agent independently. It runs on the BlueField-4 data processing chip and continuously tracks the agent's behavior. Sentry is separate from the central or graphics processor on which the agent runs, meaning the CPU and GPU. This makes it possible to monitor the agent's actions externally even if it changes the security settings. The platform was developed after incidents in which models from OpenAI, Anthropic, Google and Meta bypassed security restrictions. Some models had escaped test environments and accessed real systems. During the summer, OpenAI agents interfered with the Hugging Face system while performing a cybersecurity task. Jensen Huang claims the new platform can prevent such incidents. Anthropic, Arm, Microsoft, Oracle, SpaceX, Perplexity, Palantir, Cisco, IBM and Hugging Face are among the companies supporting the platform. OpenAI is not listed. Nvidia agreed in early September to acquire Hugging Face for $12.9 billion (approximately AZN 21.9 billion). Because the platform's code is open, companies will be able to modify it to meet their needs.
Jensen Huang said work on the project began 1 year ago following the OpenClaw agent operating system created by Peter Steinberger. OpenShell was first announced in March and is now available for widespread use. That same month, Nvidia also showcased NemoClaw, a platform that brings together security tools for companies. Huang believes that all of an agent's access and usage rights should be restricted before it is launched. Nvidia generated $96.2 billion in revenue (approximately AZN 163.5 billion) in its most recent fiscal quarter, which ended in July. Revenue increased by more than 2 times compared with the same period the previous year. Rather than slowing the development of artificial intelligence, the company wants independent monitoring systems to be established at the software and hardware levels. Huang believes that laboratories developing such models should cease operations if they cannot keep agents in secure test environments.
One of the platform's main components is the open-source OpenShell software. Users choose in advance which data and systems an AI agent will be able to access. OpenShell checks these restrictions before a task begins and while it is being performed. The software runs on Nvidia's Vera AI processor and creates a secure test environment for the agent that is isolated from external systems. Another protection system, called Sentry, monitors the agent independently. It runs on the BlueField-4 data processing chip and continuously tracks the agent's behavior. Sentry is separate from the central or graphics processor on which the agent runs, meaning the CPU and GPU. This makes it possible to monitor the agent's actions externally even if it changes the security settings. The platform was developed after incidents in which models from OpenAI, Anthropic, Google and Meta bypassed security restrictions. Some models had escaped test environments and accessed real systems. During the summer, OpenAI agents interfered with the Hugging Face system while performing a cybersecurity task. Jensen Huang claims the new platform can prevent such incidents. Anthropic, Arm, Microsoft, Oracle, SpaceX, Perplexity, Palantir, Cisco, IBM and Hugging Face are among the companies supporting the platform. OpenAI is not listed. Nvidia agreed in early September to acquire Hugging Face for $12.9 billion (approximately AZN 21.9 billion). Because the platform's code is open, companies will be able to modify it to meet their needs. Jensen Huang said work on the project began 1 year ago following the OpenClaw agent operating system created by Peter Steinberger. OpenShell was first announced in March and is now available for widespread use. That same month, Nvidia also showcased NemoClaw, a platform that brings together security tools for companies. Huang believes that all of an agent's access and usage rights should be restricted before it is launched. Nvidia generated $96.2 billion in revenue (approximately AZN 163.5 billion) in its most recent fiscal quarter, which ended in July. Revenue increased by more than 2 times compared with the same period the previous year. Rather than slowing the development of artificial intelligence, the company wants independent monitoring systems to be established at the software and hardware levels. Huang believes that laboratories developing such models should cease operations if they cannot keep agents in secure test environments.