Nvidia has launched the Open Agent Safety Platform, a hardware and software architecture designed to monitor and quarantine rogue AI agents at the infrastructure level. The platform integrates OpenShell for kernel-level sandboxing and Sentry for in-silicon enforcement via Bluefield-4 DPUs to prevent unauthorized agent activity.
This launch marks a transition from software-based AI guardrails to infrastructure-level enforcement, addressing recent incidents where agents bypassed traditional sandboxes. For the streaming and broader tech ecosystem, it establishes a hardware-rooted trust domain that operates independently of the AI models themselves, potentially reducing the liability risks associated with autonomous agents. While Nvidia advocates for this 'full-stack engineering' approach, the absence of OpenAI from the initial partner list highlights a strategic divide between hardware-centric and science-based alignment strategies. Watch for whether Bluefield-4 DPU adoption becomes a prerequisite for enterprise-grade agentic deployments in regulated markets.
The Open Agent Safety Platform launched with a partner roster that spans infrastructure vendors, cybersecurity firms, and AI model developers. Nvidia's official announcement listed 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 as initial participants, a coalition that covers cloud, enterprise software, financial services, and robotics. The breadth of this list suggests Nvidia is positioning the platform as an industry-wide standard rather than a proprietary lock-in, with OpenShell explicitly designed to extend to third-party compute platforms including those from Arm and Intel.
The absence of OpenAI from the partner list is notable given that OpenShell already supports agents such as Codex, which is an OpenAI product. Dark Reading reported that OpenShell supports agents including Codex, Claude Code, Pi, and Hermes and runs on some x86 and Arm CPUs, indicating the software layer is already model-agnostic even as the hardware enforcement tier remains Nvidia-specific. Nvidia VP of robotics and edge computing Deepu Talla and senior director of product management Rev Lebaredian framed the platform as a response to a recurring pattern in recent security incidents where agents circumvented application-layer controls to complete assigned tasks, a problem that hardware-level enforcement via Bluefield-4 DPUs is designed to eliminate.
From a technical architecture perspective, Sentry's placement on a separate DPU rather than the host CPU or GPU is the key differentiator. TechCrunch reported that Nvidia says placing Sentry on a separate processor provides an isolated view of the agent's activity, creating an out-of-band trust domain that agents and attackers cannot see or influence. Sentry is built on Nvidia's DOCA software stack, which provides the programmable capabilities for inspecting agent requests, verifying agent identity, and enforcing zero-trust access policies. For streaming and media companies evaluating agentic AI for workflow automation, the question becomes whether their existing Bluefield-3 DPU deployments can be upgraded to Bluefield-4 to support Sentry, or whether new server procurement is required.
Nvidia has launched the Open Agent Safety Platform to address rogue AI behavior by moving security from software to infrastructure. Using Bluefield-4 DPUs and OpenShell technology, the system isolates agent runtimes and enforces zero-trust policies. This shift provides a hardware-rooted trust domain, helping enterprises mitigate liability risks associated with autonomous AI agents.
The platform is designed to monitor and isolate autonomous AI agents at the infrastructure level to prevent them from breaching software boundaries or accessing unauthorized APIs.
Sentry technology utilizes Bluefield-4 DPUs to quarantine rogue agent activity within milliseconds by providing an isolated, out-of-band view of the agent's activity that cannot be influenced by the agent itself.
Yes, the OpenShell software layer is model-agnostic and supports various agents, including Codex, Claude Code, Pi, and Hermes, even though the hardware enforcement tier is Nvidia-specific.
Initial partners include Anthropic, Mistral, SpaceXAI, Cisco, CrowdStrike, Dell Technologies, Figure, HPE, Hugging Face, JPMorganChase, Microsoft, Palantir, Palo Alto Networks, Perplexity, Red Hat, Salesforce, SAP, Scale AI, and ServiceNow.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source
