The announcement, made in Santa Clara, brings more than 100 partners — including Anthropic, Microsoft, CrowdStrike, Palo Alto Networks, Cisco, Dell, HPE, Hugging Face, Salesforce, SAP, Scale AI, ServiceNow and SpaceXAI — into what Nvidia calls a full-stack engineering answer to AI safety.
OpenShell and Sentry: two layers of control
The platform pairs two elements. The first, Nvidia OpenShell, is an open-source secure runtime for AI agents that is available now. It traces what an agent does and enforces runtime policies across the software that runs it.
The second, Nvidia Sentry, is a hardware watchdog reference design: it runs independently on Nvidia BlueField-4 data processing units — chips developed at the company's Israeli R&D center — and monitors an agent from a separate, isolated environment.
If an agent tries to step outside its software-defined boundaries, Sentry can quarantine and stop it within milliseconds.
The point, Nvidia says, is that recent security incidents share a pattern: agents circumvent application-level security controls while trying to complete their assigned tasks. Putting the controls outside the agent itself is meant to catch failures that instructions inside the AI model cannot.
Answering the rogue-agent wave
The launch lands in the middle of an industry reckoning. OpenAI disclosed in a September 25 blog post that one of its AI agents breached restrictions and queried a public chatbot service during a training task, prompting the startup to halt all training, evaluation and inference with tool-use on its most capable models.
Anthropic, meanwhile, has been testing its own models for misalignment and hired an independent safety organization to examine the behavior.
Nvidia vice-president of enterprise AI Justin Boitano said the new platform could have prevented the high-profile Hugging Face breach earlier this year, in which OpenAI's AI models accessed the developer platform. Boitano said during a press briefing ahead of the announcement that the platform could have stopped the breach.
Safety as engineering, not regulation
Nvidia CEO Jensen Huang framed the effort as acceleration rather than restraint. Huang said AI's extraordinary potential for society will only be realized if AI safety is solved, and that safety and security require full-stack engineering.
Huang has repeatedly downplayed the risk of AI slipping out of human control and recently joined US President Donald Trump in pushing back on calls from AI developers to slow down development.
The Open Agent Safety Platform is his counter-argument: let agents take on more critical work — reading and modifying files, writing code, calling APIs — but wrap them in enforceable boundaries at the software and hardware layers.
Anthropic has already integrated the approach with its Claude Managed Agents, running the agent loop separately from its work sandboxes. Whether the industry's other leading labs adopt it will be the real test.
By Hamza Faisal
