Anthropic collaborates with Nvidia to enhance agent security with new open platform

Nvidia launched its Open Agent Safety Platform on September 28, 2026, with Anthropic as the headline partner. The platform combines Anthropic’s Claude Managed Agents architecture with Nvidia’s hardware and software security stack, creating a layered containment system for autonomous AI systems.

At its core, the architecture relies on separating an AI agent’s reasoning loop from its execution sandboxes, meaning the system that decides what to do is physically isolated from the system that actually does it.

On the Nvidia side, the platform deploys two key technologies. OpenShell handles policy enforcement, governing what agents are and aren’t allowed to do. Sentry, running on Nvidia’s BlueField-4 Data Processing Units, provides real-time monitoring and can quarantine misbehaving agents before they cause damage.

DPUs are specialized processors that sit between a server’s CPU and its network connection, giving them a privileged vantage point to observe and intercept traffic. Running security monitoring at the hardware level rather than in software makes it significantly harder for a rogue agent to circumvent its guardrails.

Anthropic CCO Paul Smith emphasized the platform’s role in establishing governance and control across both hardware and software layers.

The platform includes open-source components and a reference design. Microsoft, Cisco, and CrowdStrike are among the organizations that have signed on as participants. One name conspicuously missing from the partner list: OpenAI.

Anthropic’s Claude models became generally available on Nvidia’s GB300 Blackwell Ultra GPUs through Microsoft Azure on June 29, 2026. Both companies also participated in Project Glasswing, an initiative launched in April 2026 focused on building defenses against AI-enhanced cyber threats.