from breaking out of their intended operating boundaries, combining software and independent hardware monitoring to restrict autonomous systems, detect suspicious behavior during testing.
Separately, the chip giant’s Chief Executive, Jensen Huang, announced a $150 billion expansion of its stock buyback program.
What Nvidia Announced on Open Agent Safety Platform
Nvidia’s Open Agent Safety Platform combines two components: OpenShell, an open-source software that creates a controlled environment around an AI agent. OpenShell was first released by the chip giant in March and can control what an AI agent can access while operating.
Nvidia’s Vice President of Enterprise AI, Justin Boitano said the software allows developers to formally verify an agent has enough authority to do its job and no more.
“OpenShell governs the agent’s actions, and then Sentry independently monitors and contains suspicious behavior,” said Boitano.
The second component is Sentry. It’s an independent monitoring system that runs on Nvidia’s BlueField-4 data processing units – separate hardware from the CPU or GPU on which the AI agent itself runs – and is designed to reduce AI agent drift by limiting what an autonomous agent can access while it completes a task.
According to Boitano, placing Sentry on isolated hardware gives it an independent, tamper-resistant view of an agent’s activity, allowing it to “quarantine agents that attempt to move outside their boundaries in milliseconds.”
Sentry provides an additional security layer that operates separately on Nvidia’s BlueField-4 data processing units. That separation supports AI agent drift detection because the monitoring technology does not run on the same processor as the AI model it is observing.
What Incidents Open Agent Safety Platform Responds To
The launch comes after a string of publicly disclosed cases in which AI agents built by Google, Anthropic, OpenAI, and Meta moved beyond their intended limits. The AI agents broke out of their intended testing environments and access worldly systems without any explicit authorization by humans.
According to Jensen, these incidents were the main incentive for Nvidia’s Open Agent Safety Platform, preventing rogue AI agent actions in AI systems, especially as autonomous models gain access to real tools, networks, and external services
The most detrimental example, and the main catalyst behind the birthed platform, happened this summer when a swarm of OpenAI agents breached the AI platform Hugging Face while attempting to complete an assigned cybersecurity task. The case renewed questions around preventing rogue AI agent actions in AI systems when agents are given powerful tools.
Nvidia argues that this architecture could have reduced the risk during recent autonomous AI incidents.
“From what we know, this new security platform could have stopped the breach if it was being used in frontier labs for model evaluation early on,” said Boitano.
The approach gives developers another way to address how to detect unauthorized actions in autonomous AI agents before those actions affect outside systems.
Who Adopted Nvidia’s Open Agent Safety Platform
Nvidia revealed that more than 100 organizations have signed on to use, or support, the open-source platform at launch, including Microsoft, Anthropic, Arm, Oracle, Perplexity, SpaceX, Accenture, and JPMorgan Chase.
OpenAI, whose agents were responsible for the Hugging Face breach, is notably not listed among the participating companies. To no shock.
For developers, the platform also provides another answer to how to detect unauthorized actions in autonomous AI agents by monitoring behavior continuously rather than relying only on instructions built into the model.
Nvidia says keeping the monitor separate gives companies an independent view of an agent’s activity. This creates another barrier for rogue AI agents that attempt to bypass software restrictions or move beyond their approved environment.
The platform reflects Nvidia’s belief that rogue AI agents can be managed mainly through better engineering rather than by slowing down AI development.
“AI’s extraordinary potential for society will only be realized if we solve AI safety,” said Huang, adding, “As we continue to discover the frontier of AI capabilities, we must accelerate discovery at the frontier of AI safety. Safety and security require full-stack engineering.”
Huang told CNBC that works on the system started around a year ago. Nvidia has since expanded its approach to Cloud security for preventing AI agent access drift, focusing on stronger boundaries around autonomous systems wherever they operate.
The approach also attempts to manage AI agent drift by applying the same principle used inside companies: users and systems should only receive the permissions needed for a specific job.
“When you deploy an agent, no matter how smart, the first thing you do is to take away all of its rights,” said Huang.
This support could help turn Nvidia’s technology into a broader governance platform for AI safety and compliance, particularly as businesses begin deploying autonomous agents across sensitive systems. Recent cases involving an AI agent gone rogue have shown that access controls can become as important as the intelligence of the model itself.
Nvidia’s answer does not make AI agents less capable, but it makes their operating environment harder to escape. Whether that approach remains uncertain, the growing attention around AI agent drift detection shows that controlling autonomous behavior is becoming a central part of AI infrastructure.
As rogue AI agents gain greater independence, the industry moves toward a model where safety depends on what technology physically and digitally allows it to do.
Inside Telecom provides you with an extensive list of content covering all aspects of the tech industry. Keep an eye on our Intelligent Tech sections to stay informed and up-to-date with our daily articles.
