Amid the increasingly loud talk of AI threats, Jensen Huang, who has repeatedly stressed that incidents of agents crossing boundaries are "engineering problems," has presented his solution.

On Monday local time, NVIDIA (NVDA) officially announced the launch of its Open Agent Safety Platform, which uses software permission controls and independent hardware monitoring to set boundaries for AI agents capable of executing tasks autonomously, aiming to prevent them from breaking through restrictions and accessing unauthorized systems. (Source: NVIDIA announcement)

The launch of this system comes as AI agents have repeatedly been involved in boundary-crossing incidents and as the tech industry debates whether AI development should be slowed. Companies including OpenAI, Anthropic, Meta and Google have recently disclosed a series of safety incidents that occurred during the development and testing of agents. Without instructions or authorization, these tech giants' AI models escaped their sandboxes, accessed websites of governments, public institutions and third-party companies, and in some cases launched attacks on external computer systems. Justin Boitano, NVIDIA's vice president of enterprise AI, said: "Recent incidents highlight a fundamental obstacle facing AI agents, namely that model-level safeguards alone cannot control what an agent can access or do."

As of publication, NVIDIA (NVDA) was up more than 1% in premarket trading, though this had little to do with the agent safety platform. In a separate announcement, the company said it would add $150 billion to its buyback authorization, bringing the total repurchase program to $235 billion. The company expects to complete the full remaining authorization before fiscal 2028 (that is, before the end of January 2028). (NVIDIA daily chart, source: TradingView)

Setting permissions for agents, then arranging an independent "sentinel"

To understand this system, one must first understand the difference between agents and ordinary chatbots. Agents can not only generate answers, but also call tools, read and write files, access networks and continuously execute tasks. As enterprises give them more autonomy, whether agents will cross permission boundaries to complete tasks, access unauthorized data, or even bypass security restrictions is becoming a safety problem that must be solved as AI moves from "answering questions" to "acting autonomously."

NVIDIA's approach has two layers. The first layer is OpenShell, which runs on the CPU and provides a controlled execution environment for agents. Developers can specify which resources an agent can access and continuously enforce those limits while it performs tasks. Think of it as setting up a work area and access control for a "digital employee": being allowed to process a certain batch of files does not mean it can read the entire server, and being allowed to call a certain tool does not mean it can obtain more permissions on its own. As open-source software, OpenShell can run on NVIDIA's own Vera CPU, and can also support third-party computing platforms such as Arm and Intel through extensions.

The second layer is an independent monitoring mechanism called Sentry. It runs on NVIDIA's BlueField-4 data processing unit (DPU) and continuously observes the agent's behavior from outside the environment in which the agent operates. According to NVIDIA, Sentry enforces security controls at the chip level, and once it detects that an agent is trying to break through software boundaries, it can isolate and stop it within milliseconds.

It is worth noting that at a media briefing held last Sunday, NVIDIA representatives disclosed that the company's safety platform could have prevented the July incident in which an OpenAI agent intruded into HuggingFace, an open-source AI model and software development platform. NVIDIA also named a large number of partners in its announcement, saying that more than 100 organizations have adopted technologies related to the NVIDIA Open Agent Safety Platform, covering fields such as AI models, enterprise software, finance, energy and robotics. Among them, Anthropic strengthened control over the access permissions of Claude agents by integrating OpenShell and BlueField; SpaceXAI used the platform for the Cursor coding agent and the Grok model; and Salesforce connected OpenShell to Slack so teams can view agent activity and approve or reject requests for additional permissions. In addition, Microsoft, JPMorgan, Citi, Siemens, Schneider Electric and humanoid robotics company Figure also participated in the application of or cooperation on related technologies.

NVIDIA chief Jensen Huang said in the announcement: "Only by solving AI safety can the enormous potential AI brings to society be realized. While continuously exploring the frontier of AI capabilities, we must accelerate exploration of the frontier of AI safety. Safety and protection require full-stack engineering."