NVIDIA moves agent safety enforcement into silicon

NVIDIA announced the Open Agent Safety Platform, a stack that treats agent security as an infrastructure problem rather than a prompting problem. The platform pairs two pieces: OpenShell, an open source secure runtime for AI agents, and Sentry, a reference system design that adds a hardware watchdog layer.

The reasoning behind it is blunt in NVIDIA’s own framing. Recent security incidents share a pattern: the agent found a way around application layer security controls in order to finish its task. OpenShell’s answer is to place the boundary outside the model and the agent harness, so the enforcement point is a separate process that traces every action and applies policy. NVIDIA reports minimal performance overhead on Vera, described as the first CPU purpose built for agentic workloads, and says the design extends to third party compute including Arm and Intel parts.

The hardware layer

Sentry runs as an independent watchdog on BlueField-4 DPUs, able to quarantine and stop a running agent in milliseconds. The point is that it is invisible to the agent itself, so an agent has no path to detect or negotiate with it. It is built on NVIDIA DOCA, which provides the attested telemetry, agent identity verification and zero trust data policies underneath.

Named organizations working with the platform, as listed by NVIDIA, include Anthropic, Cisco, CrowdStrike, Dell Technologies, Figure, HPE, Hugging Face, JPMorganChase, Microsoft, Palantir, Palo Alto Networks, Perplexity, Red Hat, Salesforce, SAP, Scale AI, ServiceNow and SpaceXAI, alongside a claim of over 100 organizations working with the technology. NVIDIA also launched the Open Secure AI Alliance with more than 120 member organizations under Linux Foundation governance, which is the piece that matters if this is ever going to be more than a vendor stack.

Concrete integrations, not just a logo wall

A launch post full of partner logos usually means nothing was built, so the specific descriptions carry the weight here:

• Anthropic runs the agent loop for Claude Managed Agents on a separate server from the sandboxes where work executes, integrated with OpenShell and BlueField.
• Salesforce wired OpenShell into Slack so employees can watch agent activity and approve requests.
• SAP is embedding OpenShell in Joule Studio as an enterprise control point for agents interacting with SAP business systems.
• SpaceXAI uses the stack for its Cursor coding agents and Grok.

OpenShell is broadly available today and published under an open source license, with the Sentry design offered as a reference rather than a turnkey product. The honest read is that NVIDIA has made a credible architectural argument and shipped the open source half of it, while the hardware half is still a design you implement yourself. Adoption among the named partners is what will tell you whether that gap matters.

(Original source: NVIDIA)