Nvidia Launches New Security System to Stop Rogue AI Agents

Nvidia has introduced a new security platform designed to keep autonomous AI agents within defined limits as companies increasingly use them to perform tasks with limited human intervention. The Nvidia Open Agent Safety Platform combines software-based controls with a separate hardware monitoring layer. The company says the system is intended to monitor AI agents continuously and intervene when they attempt to operate outside their approved boundaries.

At the centre of the platform are two technologies, OpenShell and Sentry. OpenShell is an open-source secure runtime that places AI agents inside sandboxed environments and allows organizations to specify which files, networks, tools, processes and credentials an agent can access. Sentry provides an additional layer of protection. The technology runs on Nvidia's BlueField-4 data processing units (DPUs) and operates separately from the AI agent's software environment. Nvidia says Sentry can detect policy violations and quarantine an agent within milliseconds if it attempts to move beyond its permitted operating area.

The announcement comes as the AI industry faces growing concerns about autonomous systems behaving in unexpected ways. Unlike conventional chatbots, AI agents can interact with external tools, access data, execute code and perform tasks over extended periods. That greater autonomy also creates new security challenges.

Nvidia said recent incidents involving AI agents demonstrated the limitations of relying only on application-level safeguards. The company argues that security controls should exist outside the AI model itself so that an agent cannot simply bypass the restrictions it is supposed to follow. The company has also said its new technology could have prevented a recent security incident involving AI platform Hugging Face. That assessment is Nvidia's own claim; Reuters reported that the company presented the new tools as a response to broader concerns surrounding autonomous AI systems.

Nvidia's platform is being developed with companies across the AI and technology industries. Participants include Anthropic, Microsoft, Cisco, CrowdStrike, IBM, SAP, Salesforce, Palantir, Hugging Face and others, according to Nvidia. The company said more than 100 organizations are working with its platform technologies.

The approach reflects a broader shift in AI security. As autonomous agents receive greater access to corporate systems and digital tools, organizations increasingly need controls that can restrict permissions, monitor activity and stop an agent when its behaviour moves outside an approved scope. Nvidia's new platform is aimed at providing that additional layer of control, potentially making autonomous AI systems easier for businesses to deploy while retaining a mechanism for human oversight and intervention.

Disclaimer: This image is taken from NDTV.