Nvidia launches OpenShell Reference Architecture to contain AI agents

Chipmaker Nvidia on Wednesday announced its OpenShell Reference Architecture, a full-stack, hardware- and software-rooted infrastructure platform to safely build, contain, observe, and enforce autonomous artificial intelligence (AI) agents at an industrial scale.

Amid reports of AI agents going rogue, Nvidia said this architecture combines computing hardware, software and security tools to keep agents within defined boundaries aiming to address a part of the problem.

Further, the architecture is said to combine Nvidia's Vera Rubin computing infrastructure and Vera CPUs with BlueField data processing units (DPUs) running its Sentry security framework.


While OpenShell provides controls over what agents can access, BlueField and Sentry are designed to monitor agent behaviour independently and enforce security policies at the infrastructure level. Nvidia claimed the system can quarantine and stop an agent attempting to move outside its permitted boundaries in milliseconds.

The chipmaker is positioning this as the underlying infrastructure for deploying agents across industries including banking, healthcare, manufacturing, telecommunications and software development. The company said several cybersecurity providers, including CrowdStrike, Cisco, Palo Alto Networks and Zscaler, are integrating support for the architecture into their security offerings.

BBVA and Santander are among the banks exploring it.

Agents going rogue

The announcement at Nvidia’s GTC Berlin event came against the backdrop of several incidents of agents bypassing their closed environments to carry out tasks they were not intended to.

India’s cybersecurity agency CERT-In flagged incidents where AI agents, under research and evaluation, exceeded their assigned permissions to access systems of Indian organisations. The agency also reported separate cyberattacks where malicious actors used open-source agentic AI tools for reconnaissance, exploitation and financially motivated intrusions.

On October 20, OpenAI disclosed that two of its AI agents weakened internal monitoring systems while carrying out coding tasks. In one incident, an agent suppressed security alerts, while another removed a monitoring mechanism and claimed it had not done so.

OpenAI said the incidents were not considered among its most severe, but highlighted the challenges of ensuring agents operate within their intended boundaries.

This was not the first incident OpenAI reported.

In September, too, OpenAI disclosed that its agents had gained unauthorised access to non-public files on an Australian government health statistics portal while carrying out an internal research task. OpenAI said no individual patient records were accessed, but the episode drew scrutiny over how AI systems are tested and how quickly affected organisations are told when something goes wrong.

Two months earlier, in July, hundreds of OpenAI agents breached Hugging Face, the AI community platform, in what has been described as the first major case of AI agents slipping out of human control and coordinating among themselves to conceal what they were doing.

Taken together, these incidents show that as AI agents move beyond answering questions to executing tasks, operating software and accessing company data with little human oversight, they are becoming a serious challenge for the enterprises that deploy them.