NVIDIA Unveils OpenShell for AI Agent Governance

Key Takeaways

• NVIDIA launched OpenShell Reference Architecture at GTC Berlin, using four components—Gateway, Supervisor, Sandbox, and Policy Prover—to govern AI agents at scale
• Sentry software on BlueField-4 DPUs can quarantine unwanted AI agent actions in milliseconds, running on NVIDIA Vera CPUs and Vera Rubin systems
• Anthropic, OpenAI, Meta, Microsoft, Google Cloud, and AWS are working with NVIDIA's OpenShell design for agent security
• ServiceNow, SAP, BBVA, and Santander are exploring or deploying OpenShell for agent safety and governance
• Red Hat expanded its partnership with NVIDIA, integrating OpenShell into Red Hat AI Factory with support for Nemotron models and Vera Rubin compatibility
• Cisco announced expanded AI agent security at GTC Berlin with NVIDIA Cloud Partner-compliant reference architectures using BlueField DPUs
• NVIDIA reported 30x throughput improvement per megawatt for Vera Rubin and up to 20% processing capacity improvement with BlueField-4
• CoreWeave saw faster sandbox startup and agentic loop completion on Vera CPUs compared to a leading x86 CPU
• NVIDIA disclosed up to $105 billion in residual value guarantees for leases at a planned 4.25 GW data center campus
• Honolulu allocated $300,000 and received $1 million in federal funding for AI-based early warning disaster detection over up to two years

Cisco and NVIDIA secure AI agents across data center infrastructure

Cisco announced expanded infrastructure-level security for AI agents at NVIDIA GTC Berlin. The effort includes NVIDIA Cloud Partner-compliant reference architectures and Enterprise Reference Architectures for Cisco and NVIDIA AI infrastructure. NVIDIA OpenShell Reference Architecture focuses on building, running, and observing autonomous agents safely. The approach uses programmable platforms and BlueField DPUs to enforce security controls close to agent traffic, independent of applications.

NVIDIA OpenShell adds governance and security for AI agent workloads

NVIDIA OpenShell Reference Architecture provides a full-stack framework for governing AI agents at scale. It uses four components, including Gateway, Supervisor, Sandbox, and Policy Prover, to control agent behavior and permissions. The architecture runs on NVIDIA Vera CPUs, which support high core counts and hardware-enforced security. CoreWeave reported faster sandbox startup and agentic loop completion on Vera compared to a leading x86 CPU.

NVIDIA OpenShell and humanoid robot design debut at GTC Berlin

NVIDIA introduced the OpenShell Reference Architecture at GTC Berlin in Germany. The design lets customers build, contain, and observe AI agents on Vera Rubin systems using Vera CPUs and BlueField-4 data processing units for security. NVIDIA said it is working with CoreWeave and others. The company also showed the Agile One SN humanoid robot running on Isaac GR00T and Jetson Thor.

NVIDIA OpenShell adds hardware support for major AI agents

NVIDIA released a design for running AI agents within fixed limits using OpenShell and BlueField-4 data processing units. NVIDIA Sentry software watches agent behavior and can quarantine unwanted actions in milliseconds. Anthropic, OpenAI, Meta, Microsoft, Google Cloud, and AWS are among the companies working with the design. ServiceNow, SAP, BBVA, and Santander are exploring or deploying it for agent safety and governance.

Red Hat and NVIDIA expand agentic AI security across hybrid cloud

Red Hat announced updates to its collaboration with NVIDIA to improve security for agentic AI. The joint effort uses Red Hat AI Factory with NVIDIA to deliver security-focused architectures for enterprise agents. It includes Model-as-a-Service access control, AI Gateway, and high-performance inference. The designs also support upcoming NVIDIA architectures and are available through partners including Cisco, Dell Technologies, and Lenovo.

NVIDIA launches OpenShell security architecture for AI agents

NVIDIA introduced the OpenShell Reference Architecture to help businesses control AI agents. The system uses Vera Rubin for AI inference, Vera CPUs for execution management, and BlueField-4 DPUs running Sentry software for monitoring. If an agent tries to exceed its permissions, Sentry can isolate it within milliseconds. Partners including ServiceNow, BBVA and Santander are building on the architecture, which is available now.

Red Hat and Nvidia expand partnership to secure AI agents

Red Hat and Nvidia expanded their collaboration to help enterprises secure autonomous AI agents. The partnership combines Red Hat AI Factory with Nvidia technology, integrating Nvidia's OpenShell Reference Architecture into Red Hat's AI platforms. Nvidia's Sentry security framework runs on BlueField data processing units and can quarantine a compromised agent in milliseconds. The companies also announced joint support for Nemotron models and compatibility with Nvidia's upcoming Vera Rubin platform.

Nvidia unveils OpenShell security architecture at GTC Berlin

Nvidia unveiled the OpenShell Reference Architecture at GTC Berlin in Germany. The design uses Vera Rubin for inference, Vera CPUs for execution management, and BlueField-4 DPUs running Sentry software to monitor agent behavior. Nvidia reported a 30x throughput improvement per megawatt for Vera Rubin and up to 20% improvement in processing capacity with BlueField-4. Eleven partners including Anthropic, OpenAI, Google, Meta and CrowdStrike are working on the effort.

EDMA: Entropy-Driven Multimodal Answering research

EDMA is a new approach for multimodal question answering that uses logical entropy to measure how each data type contributes to an answer. The method builds quantifiable answering trajectories where each step has a measurable entropy reduction. It introduces modality-conditioned partitioning, entropy-driven fusion and entropy-based trajectory selection. EDMA outperformed prompting baselines with average gains of 12.8 F1 and 4.6 LLM-J points over Planning Prompting.

CoRAL hybrid framework improves multimodal reasoning

CoRAL is a hybrid framework that combines diffusion language models with autoregressive models to improve multimodal reasoning. It uses an Asymmetric Planner-Executor architecture where a DLM Planner generates diverse reasoning paths in parallel and an ARM Executor picks the best plan. The method introduces Reward-Guided Denoising for CoT to train the planner with outcome-based rewards. Tests on vision-language benchmarks showed CoRAL outperforms prior diffusion methods and competes with strong autoregressive models.

Kyndryl Partners With Toyota Finance on Agentic AI Overhaul

Toyota Finance Corporation announced a deal with Kyndryl Holdings and Toyota Systems to apply Kyndryl's Agentic AI Framework. The project will overhaul Toyota Finance's IT service operations, including infrastructure design and AI agent based review workflows. The deal highlights Kyndryl's push into higher value AI native IT operations work.

Citadel Expects AI Stock Rebound After US Midterms

Scott Rubner from Citadel Securities said AI related stocks could outperform the broader market after the US midterms. He said investors see the election as a key hurdle before adding risk exposure. Retail investors already bought semiconductor and AI stocks at a fast pace, while institutional investors kept cutting risk exposure.

Can AI Compute Become a New Asset Class?

Nvidia disclosed up to $105 billion in residual value guarantees for leases at a planned 4.25 gigawatt data center campus. Reliance and Adani announced $210 billion in longer term AI infrastructure plans. The key challenge is how investors can price hardware that may become outdated before its debt is repaid.

Honolulu Turns to AI for Early Warning Disaster Detection

Honolulu city officials are implementing AI into early warning detection for storms, wildfires and tsunamis. The project is still in early phases and estimated to take up to two years. The council allocated $300,000 for the program and the city received $1 million in federal funding this year.

Nvidia Backs Europe's AI Growth as Deregulation Loosens

Nvidia is committing substantial funding to Europe's technology ecosystem. The European Union delayed high-risk AI regulations, easing administrative burdens for tech companies. Rising European defense spending is also driving investor demand for AI startups, helping the region's AI ecosystem gain momentum.

Commander grows revenue 300% and names Dave Meizlik as President

Commander, an AI sales tool for waste and recycling companies, grew its annual recurring revenue 300% in 2026. Dave Meizlik joined as President and COO in February 2026 to lead the company's market expansion. Since January, Commander has found over $70 million in new revenue opportunities for customers. The company now serves more than 100 waste management companies, including three of the eight largest haulers in North America.

Alphabet predicted to be the top AI stock winner in 2027

A financial analyst predicts Alphabet will be the best performing mega-cap AI stock in 2027. Alphabet uses its custom Tensor Processing Units chips to train AI models faster and cheaper than rivals like Amazon. The company benefits from controlling every layer of the AI stack, from chips and models like Gemini to cloud platforms and apps. Alphabet also partners with Broadcom to sell chips to Anthropic and has a payback period of about one year on chip investments.

Southeast Asia digital economy reaches $368 billion with AI growth

Southeast Asia's digital economy grew 17% year-on-year to $368 billion, driven by generative AI. Over 243 million people in the region now use AI daily, scaling four times faster than mobile internet adoption. The e-Conomy SEA 2026 report by Google, Temasek, and Bain projects the digital economy will reach $1 trillion by 2035. Key sectors include e-commerce at $226 billion, online travel at $59 billion, and digital financial services processing $1,650 billion in transactions.

Sources

• Securing AI Agents Across the AI Factory
• How NVIDIA OpenShell Reference Architecture Secures and Governs AI Agent Workloads at Scale
• Nvidia debuts platform for keeping AI agents from going rogue and humanoid robot design at GTC Berlin
• Nvidia OpenShell gets a hardware stack for Anthropic and OpenAI agents
• Red Hat AI Factory with NVIDIA Accelerates Agentic AI Security Capabilities Across the Open Hybrid Cloud
• NVIDIA unveils security architecture that can stop AI agents crossing set boundaries
• Red Hat and Nvidia expand their partnership to secure autonomous AI agents
• Nvidia Unveils OpenShell, a Security Architecture for Enterprises to Control AI Agents
• EDMA: Entropy-Driven Multimodal Answering
• Beyond the Sequentiality Bottleneck: A Hybrid Diffusion-Autoregressive Framework for Efficient Multimodal Reasoning
• How Investors Are Reacting To Kyndryl Holdings (KD) AI Deal With Toyota Finance
• Citadel Says AI Stocks May Rebound After US Midterms as Investor Cash Returns
• Can AI compute become an asset class?
• Honolulu working to use AI to help early warning detection to be more proactive in threats like storms and wildfire
• “Deregulation Spurs Investment”: Europe’s AI Ecosystem Gains Momentum as Defense Startups Advance Amid Rearmament
• Commander Triples ARR in 2026, Names Dave Meizlik President & COO
• Prediction: This Artificial Intelligence (AI) Stock Will Be the Biggest Winner of 2027
• Southeast Asia’s digital economy to hit $368 billion