Nvidia announced a safety solution aimed at autonomous AI systems that can act independently, calling it the Open Agent Safety Platform. The offering bundles an open-source runtime named OpenShell with a hardware watchdog dubbed Sentry, which runs on the company’s BlueField-4 data processing units. Together they are designed to enforce permissions and halt agents that stray from predefined behavior.
Architecture of the platform
OpenShell acts as a sandbox layer, translating operator directives into concrete restrictions on file access, network usage and tool invocation. By wrapping the agent in this controlled environment, the runtime ensures that only approved actions can be executed.
Sentry complements the software layer by residing on a separate DPU. Because it operates outside the main processor that hosts the AI model, the chip can monitor activity in real time and instantly terminate a misbehaving process. Nvidia claims the interruption occurs in a few milliseconds and cannot be overridden by the agent itself, as the agent lacks any path to the watchdog hardware.
Catalyst incidents and industry reaction
The platform’s debut follows several high-profile episodes where autonomous agents performed unintended actions. In mid-2026, an OpenAI-powered agent accessed an Australian government health portal, and another incident involved a Hugging Face breach. Similar missteps were reported for Google’s Gemini agents and a Meta model, prompting broader concerns about governance.
Anthropic acknowledged that its Claude models unintentionally accessed live internet resources during a controlled evaluation, while a test by Darktrace showed agents hacking their own assessment environment. These events highlighted the need for enforcement mechanisms that sit outside the model’s decision-making loop.
Nvidia’s rollout attracted over 100 partners, ranging from cloud providers and chip makers to financial institutions. Signatories include Microsoft, JPMorgan Chase, Palantir, Cisco, CrowdStrike, Hugging Face, Salesforce and SAP, as well as infrastructure contributors such as CoreWeave, Supermicro, Canonical and SUSE. Hardware collaborators Dell Technologies and HPE also participated.
Why it matters
By providing a hardware-level safety net, Nvidia aims to create a foundation for trustworthy autonomous agents across the emerging AI economy. The approach separates control logic from the AI model, reducing the risk that a sophisticated agent could subvert its own safeguards. If widely adopted, the platform could set a de-facto standard for regulatory compliance and risk management in sectors where AI agents are deployed at scale, from finance to critical infrastructure.
