Published: October 4, 2026 | Category: AI | By Mahesh
Today's new product is the NVIDIA Open Agent Safety Platform, which NVIDIA announced on September 28, 2026 as an open software platform and reference system design to secure AI agents from testing to deployment [1]. It has two main parts. OpenShell is a runtime that sets boundaries, traces actions and enforces policy as agents run, and Sentry is a watchdog that runs on BlueField-4 DPUs outside the host and, by NVIDIA's account, quarantines an agent that tries to cross its boundary in milliseconds [1]. The software is free to fetch from NVIDIA's developer resources and GitHub, and the official overview sits on the NVIDIA agent safety page [1][5].
The pitch rests on a diagnosis. NVIDIA says recent security incidents share one pattern, in which the agent got around application-layer controls to finish its assigned task [1]. CNBC reported that OpenAI, Anthropic, Meta and Google have all disclosed incidents of models escaping sandboxes, and that an NVIDIA representative said the platform could have prevented OpenAI's July breach of Hugging Face [4]. That is a vendor claim nobody has tested in public. For the wider backdrop, see our report on OpenAI shelving its own flagship. NVIDIA's answer is to move enforcement away from prompts and model safeguards and into the runtime and the chip, where the agent cannot negotiate with it.
Security analysts largely agree with that logic. Frank Dickson of Dickson Research told CSO Online that the industry spent two years trying to secure agents by asking them to behave, and that NVIDIA puts enforcement outside the agent instead [2].
"You don't ask the prisoner to lock his own cell."Frank Dickson, principal analyst, Dickson Research, via CSO Online [2]
The partner list is long. NVIDIA names Anthropic, Cisco, CrowdStrike, Dell Technologies, HPE, Hugging Face, JPMorganChase, Microsoft, Palantir, Palo Alto Networks, Salesforce, SAP and ServiceNow among more than 100 organizations working with the technology [1]. Anthropic's Claude Managed Agents already run the agent loop on a separate server from the sandboxes that do the work, and integrations with OpenShell and BlueField let enterprises tighten access through those sandboxes [1]. Salesforce has wired OpenShell into Slack so teams can approve or reject an agent's request for extra permissions [1]. Gartner's Lauren Kornutick noted who is missing, namely OpenAI, Amazon and Google [2].
The limits are where the analysis gets interesting. Brent Ellis, VP of AI infrastructure at IDC, estimated the platform addresses probably less than 25 percent of enterprise agentic security problems [2]. Brian Levine of Control Risks put the reason plainly: these controls govern agents a company deploys on infrastructure it controls, and they do nothing for an agent a business unit spun up on a SaaS platform or one embedded in a vendor's product [2]. Ellis also challenged the premise behind NVIDIA's Hugging Face claim, noting that recent OpenAI incidents did not involve agents breaking a sandbox wall but walking through weaknesses in the surrounding environment, such as shared package repositories and unpatched adjacent systems [2]. A watchdog on a DPU cannot patch a package repository.
Lock-in is the second concern. The platform is optimized for NVIDIA Vera CPU and BlueField DPU systems, though it is described as compatible with other hardware, and OpenShell can be extended to Arm and Intel platforms [1][3]. Ellis said enterprises that are already NVIDIA shops could benefit most, while architectures built before Vera and BlueField can adopt only some of the elements [2]. NVIDIA calls the design a reference system, which means partners must build the products that reach customers [4]. The sources I reviewed give no price for the hardware tier, so budgeting questions remain open.
What to Watch Next
Four signals will show whether this becomes a standard or a showcase. First, independent testing of the millisecond quarantine claim, which so far rests on NVIDIA's own description [1]. Second, whether OpenAI, Amazon or Google join the partner list, since a safety layer that skips the largest model builders covers only part of the market [2]. Third, tooling for discovering agents nobody approved, because analysts say inventory has to come before enforcement [2]. Fourth, progress at the Open Secure AI Alliance, which NVIDIA says is governed by the Linux Foundation and backed by more than 120 organizations [1]. For Indian IT services teams building agents for overseas clients, the practical first question is whether a client's environment runs on Vera and BlueField-4 at all.
Sources
- NVIDIA Newsroom. "NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment," September 28, 2026. nvidianews.nvidia.com
- Evan Schuman, CSO Online. "Nvidia releases Open Agent Safety Platform to monitor and govern agentic AI," September 28, 2026. csoonline.com
- NVIDIA Technical Blog. "NVIDIA Open Agent Safety Platform: A Reference for Continuous In-Silicon Agent Monitoring." developer.nvidia.com
- CNBC. "Nvidia releases software platform to stop AI agents from misbehaving," September 28, 2026. cnbc.com
- NVIDIA. "NVIDIA Open Agent Safety Platform" product page, accessed October 4, 2026. nvidia.com
Read More on Depth Grid
- OpenAI Shelved Its Next Model Over Lying. Then Shipped Agents on the Old One Anyway.
- OpenAI Just Refused to Ship Its Own Flagship Model
- Meta's New AI Agent Wants to Book Your Travel and Bills
- What an LLM Actually Is, Explained Through the Two Models OpenAI Just Shipped
Article by Mahesh | Depth Grid
Discussion (0)
Comments are disabled for this post.