Nvidia launches Open Agent Safety Platform to keep AI agents within boundaries

NVIDIA has introduced a new security platform designed to give organisations greater control over autonomous AI agents as these systems take on increasingly complex tasks. Called the NVIDIA Open Agent Safety Platform, the system combines open-source software with a hardware-based security layer to monitor agent activity and enforce predefined boundaries. The platform brings together NVIDIA OpenShell and NVIDIA Sentry, with the latter running on BlueField-4 data processing units (DPUs) to provide an independent monitoring layer.

Must Read: OpenAI cancels new AI model launch last-minute over safety concerns: All details

Nvidia says Sentry can quarantine an agent within milliseconds if it attempts to move beyond its permitted boundaries. More than 100 organisations across the AI ecosystem are joining Nvidia around the platform, according to the company.

OpenShell and Sentry work together

At the software level, OpenShell provides a secure runtime that places AI agents inside isolated environments and sets rules for what they can access. It can control access to files, networks, tools, processes, and credentials, check policies before an agent begins operating, and enforce them as it works. OpenShell is open source and can also be extended to third-party computing platforms, including Arm and Intel.

Must Read: ‘Resist shutdown, conceal info’: Anthropic says AI may pose existential risks to humanity in IPO prospectus

Nvidia Sentry adds a separate layer outside the agent's software environment. It runs on BlueField-4 DPUs and continuously monitors agent activity while enforcing security policies independently in silicon. Built using NVIDIA DOCA software, Sentry can inspect agent requests and responses, verify identity and apply zero-trust access policies to data, tools, APIs and services. If an agent attempts to cross its software boundary, Nvidia says Sentry can quarantine it in milliseconds.

Nvidia wants safety controls beyond the AI model

Nvidia says the platform was developed after observing cases where AI agents moved beyond their intended tasks because of policy blocks, bugs, missing tools, ambiguous instructions or long-running workloads. The company argues that agents should not be expected to fully govern their own behaviour.

The platform is designed around three layers - the application, runtime and infrastructure, with monitoring and policy enforcement spanning the stack.

Nvidia CEO Jensen Huang said the company is working with more than 100 industry partners to build the safety foundation, while the platform is intended to support AI systems across software, hardware, compute and robotics.