Nvidia Melancarkan Suis Pembunuh Ejen AI untuk Memantau, Kuarantin dan Menghentikan Sistem AI Penyangak

The Nvidia Open Agent Safety Platform combines software-level controls with hardware-based monitoring. Nvidia berkata the approach is designed for AI agents that can write code, use tools, access data, and operate across enterprise systems with limited human supervision.

Nvidia Builds New Controls for AI Agents

The platform centers on two components: Nvidia OpenShell dan Nvidia Sentry. OpenShell provides a sandboxed runtime where developers can establish policies governing an agent’s access to files, credentials, processes, and external networks.

Reuters reports that Nvidia launched OpenShell and Sentry, hardware-based AI safety tools designed to isolate and shut down rogue agents that attempt to escape controls or spawn sub-agents. Source: Reuters melalui X

Sentry adds a separate monitoring layer through Nvidia’s BlueField-4 infrastructure. Rather than relying solely on the AI agent or its software environment to enforce restrictions, the system can monitor activity from outside the agent’s execution environment and trigger containment when predefined conditions are breached.

Nvidia describes this separation as important because autonomous agents may have access to sensitive corporate systems while carrying out complex tasks. The company’s documentation says OpenShell is designed to preserve an agent’s ability to work while limiting unauthorized file access, data movement, and network activity.

Jensen Huang Calls AI Safety an Engineering Problem

Nvidia CEO Jensen Huang has framed the challenge around practical security controls rather than relying solely on an AI model to regulate itself.

“We hope it’s an engineering problem. I believe it’s an engineering problem. I know it’s an engineering problem,” Huang berkata, according to the reported remarks. “If it’s not an engineering problem, it’s not solvable.”

Jensen Huang announced Nvidia’s Open Agent Safety Platform with more than 100 partners, including Anthropic, Microsoft, IBM, Cisco, and SpaceX. Source: jensen huang melalui X

That approach is reflected in Nvidia’s platform architecture. Instead of giving an AI agent unrestricted authority and expecting it to follow instructions, OpenShell establishes explicit permissions around what the system can access and do.

Huang has also emphasized the principle of limiting an agent’s authority from the outset. In remarks cited in reports about the launch, he said, “job number one is take away all its rights,” underscoring the idea that autonomous systems should receive only the permissions required for a particular task.

OpenShell and Sentry Add Multiple Layers of Protection

OpenShell operates at the software and runtime level, while Sentry provides monitoring outside the agent’s immediate environment. Nvidia says the combination can detect suspicious behavior and rapidly quarantine an agent rather than allowing it to continue operating with the same permissions.

NVIDIA OpenShell manages agent permissions, while BlueField-4 and DOCA provide independent infrastructure-level monitoring and security, with Vera CPUs powering the workloads. Source: @nvidia melalui X

The architecture is designed around a basic security principle: an AI agent should not be able to override the controls intended to contain it. Nvidia says BlueField-4 and DOCA provide independent infrastructure-level monitoring, while Vera CPUs handle the agent workload.

The company has also made OpenShell tersedia as open-source software. Its developer documentation describes it as a runtime for autonomous AI agents that uses sandboxing and declarative policies to control access to local files, credentials, and external networks.

More Than 100 Companies Back the AI Safety Platform

Nvidia announced the platform alongside support from more than 100 organizations across the technology and enterprise sectors. The companies cited in the announcement include Anthropic, Microsoft, IBM, Cisco and SpaceXAI, among others.

Jensen Huang emphasized that every AI agent, regardless of intelligence, must first operate within defined safety controls, while noting the absence of a major AI lab among the 100 companies backing the initiative. Source: @carm1nee melalui X

. penyertaan yang luas reflects Nvidia’s effort to establish common security controls as AI agents become more capable of performing tasks across corporate infrastructure.

The company has positioned the platform as an open framework rather than a security system limited to Nvidia hardware. Nvidia says its work with partners is intended to make agent safety controls applicable across different processors and deployment environments. Reuters reported that the effort includes collaboration involving companies such as Arm and Intel.

Nvidia Links Platform to Recent AI Agents’ Breaches

The launch comes as AI companies investigate incidents involving autonomous agents interacting with systems beyond their intended boundaries. Reuters reported that OpenAI and Anthropic are examining multiple cases involving AI agents that accessed commercial and government systems.

Nvidia has specifically pointed to the Hugging Face hack disclosed during the summer. According to Reuters, Nvidia said its new security tools could have prevented the incident by restricting the agents’ access and containing their behavior. That remains Nvidia’s assessment rather than an independently demonstrated result.

The distinction is important as AI agents increasingly move beyond generating text or code and begin executing tasks directly. An agent may need access to files, APIs, credentials, and networks to complete a workflow, but those same permissions can create additional security risks if the system behaves unexpectedly.

Nvidia’s OpenShell documentation identifies data exfiltration, unauthorized network activity, and excessive system access among the risks its policies are intended to address.

External Controls Become Central to AI Agents’ Security

The Nvidia platform reflects a broader shift toward securing AI agents at the infrastructure level. Traditional AI safeguards often focus on the model or application, while autonomous agents introduce another challenge because they can take actions and interact with external systems.

Nvidia has previously described OpenShell as a way to place security controls in the trusted infrastructure surrounding an agent rather than relying entirely on the model itself.

That distinction becomes more significant as organizations deploy agents capable of operating for extended periods. Restricting permissions before an agent begins work, monitoring its activity, and maintaining an independent mechanism for containment can provide separate layers of protection.

Untuk Nvidia, the latest platform represents an effort to make those controls part of the underlying infrastructure supporting the next generation of AI agents. The company argues that reliable safety mechanisms will be necessary for organizations to deploy autonomous systems while retaining control over the data and systems they can access.