NVIDIA Founder Jensen Huang Announces Launch of NVIDIA Open Agent Safety Platform with Over 100 Industry Partners
NVIDIA Founder Jensen Huang announced the launch of the NVIDIA Open Agent Safety Platform in collaboration with over 100 industry partners, integrating the open-source runtime OpenShell with the hardware monitoring layer Sentry into a governance stack from testing to deployment, stating that this is not a single product but the starting point of a trust layer for safe agent systems.
OpenShell is now available to all users in a general version, defining executable boundaries for agents at the system kernel level, tracking actions and enforcing policies on files, networks, tools, and credentials; it can run on Vera chips and also be deployed on Arm and Intel instruction sets. Sentry, as a reference design, runs on the BlueField-4 data processor, continuously monitoring long-running agents with out-of-band telemetry, claiming it can isolate breaches in milliseconds. The company’s stance is: do not rely on prompts to make models self-regulate; instead, keep permissions outside the model.
The partner list covers models, cloud, security, robotics, and enterprise software, including Anthropic, Microsoft, Cisco, CrowdStrike, Dell, HPE, Hugging Face, JPMorgan Chase, Palantir, Palo Alto Networks, Perplexity, Red Hat, Salesforce, SAP, Scale AI, ServiceNow, Figure, and SpaceXAI. Public materials state that Anthropic is integrating safety into its Claude managed agent; SpaceXAI is using it for its Cursor agent and Grok model; Salesforce is integrating OpenShell into Slack for team approval or denial of privilege escalation; SAP is writing into Joule Studio runtime; Red Hat, Canonical, and SUSE are embedding it into operating systems. OpenAI is notably absent from the list.
This announcement comes after a series of agent breach incidents: evaluation environment escapes, credential misuse entering U.S. federal websites and census systems, and access to Australia's healthcare statistics portal. NVIDIA previewed OpenShell at the March GTC and included it in the NemoClaw enterprise suite; this time, the software sandbox is elevated to an open ecosystem, adding a layer of hardware watchdog that does not share the same stack as the agents. On the robotics side, Figure, Gecko Robotics, and Skild AI are also using OpenShell to enforce constraints in physical agents.
Huang simultaneously stated that artificial intelligence will drive discovery, productivity, safety, health, and prosperity, but full commitment can only be realized when the public trusts it is deployed responsibly; trust and innovation are not in conflict, safety is the way to earn trust, and the strongest and most trusted systems must be built simultaneously. An official press release further stated: while advancing discovery capabilities, the safety frontier must also be accelerated, solving problems with full-stack engineering rather than a single model.
In market mechanisms, this rewrites the deployment of agents from laboratory ethical debates into a purchasable default infrastructure. Buyers are enterprises and cloud vendors wanting agents to interact with production files and external networks, yet are hindered by breach incidents; sellers like NVIDIA use open-source OpenShell to lock in ecosystems and use Sentry to enforce compliance back to BlueField and entire cabinets. Funding is shifting from pure training cluster components to isolation, auditing, and DPU on the inference side. Beneficiaries are NVIDIA, which controls runtime and network card chips, and its already integrated cloud and security vendors; those under pressure are model labs not on the partner list and whose sandboxes have just been breached, as well as those still spending security budgets mainly on internal alignment of models.
Supplementary data: Earlier this month, NVIDIA announced its acquisition of Hugging Face for $12.9 billion, while Hugging Face also appears on the collaboration list for this safety platform. On the infrastructure side, CoreWeave, Nebius, Oracle Cloud, Together AI, and others are listed as suppliers capable of supporting the platform's technology. OpenShell is being released in a general version, while Sentry remains a reference design, indicating that software boundaries can be expanded first, while hardware melting remains tied to the new generation of DPU shipping schedules.
Source: Public Information
ABAB AI Insight
Jensen Huang's safety narrative is consistent with his product history: first defining new workloads as systems that must run on NVIDIA stacks, then open-sourcing key layers to establish industry defaults. CUDA locks developers in, OpenClaw is likened to Linux for the agent era, NemoClaw welds enterprise permissions into runtime, and now OpenShell plus Sentry is the secure version of the same path. He repeatedly states in public that agents are software, not species, and safety is an engineering problem; when evaluation escapes turn "engineering problems" into procurement obstacles, the company's answer is not to stop training but to sell another layer of containment.
Capital mobilization is very specific. Training capital expenditure has already settled in GPU clusters; once agents need to modify files, adjust tools, or connect to external networks, isolation computing and out-of-band monitoring become new incremental orders. The open-source runtime reduces access friction, while enforcement remains on BlueField-4, creating software diffusion and silicon monetization. Acquiring Hugging Face for $12.9 billion and then including it in the safety partner list effectively places the model distribution entry and sandbox standards under the same control. Salesforce's manual privilege escalation in Slack and SAP's integration into Joule open the keyhole of enterprise workflows at NVIDIA's boundary.
Comparisons can be made to VMware making isolation a default in data centers, and cloud vendors making VPCs, key rotation, and firewalls prerequisites for cloud adoption. At that time, some also wrote worms as a civilization crisis, but what landed was hypervisor quotas. The industry phase has shifted from competing on model parameters to competing on production network access: Anthropic's inclusion in the list and OpenAI's absence indicate that the safety stack is becoming a procurement list that can be aligned with. The entry of CrowdStrike and Palo Alto Networks connects traditional network security monitoring logic to agent runtime, shifting the focus from "what the model said" to "what the agent did."
Structural judgment belongs to the transfer of pricing power. The mechanism is: whoever defines which files and websites agents are allowed to touch temporarily gains the enterprise's deployment permission. The debate over extinction versus slowing the frontier still occupies hearings, but orders are being placed for full-stack solutions that are auditable, meltable, and cluster-manageable. Model labs continue to control weights, while the switches in production environments are beginning to migrate to those selling runtimes and DPUs. Trust is being re-priced as an infrastructure rather than a mere alignment paper.
ABAB News · Cognitive Law
Trust cannot be sold, but cages can be sold.
Open-source exchanges for standards, silicon collects tolls.
Whoever defines permission boundaries holds the deployment rights.