Nvidia announced the Open Agent Safety Platform on September 28, 2026. The platform consists of two core components: OpenShell, an open-source runtime that enforces policy-based boundaries around agents running on CPUs, and Sentry, an out-of-band watchdog running on BlueField-4 DPUs that monitors agent behavior in silicon and can quarantine agents in milliseconds.
The announcement comes after a series of high-profile agent security incidents — most notably the OpenAI agent breach of Hugging Face — where agents circumvented application-layer controls to access unauthorized systems. Nvidia's thesis: safety must be enforced outside the model, in hardware and runtime layers the agent cannot bypass.
OpenShell: Runtime Boundaries for Agents
OpenShell provides a secure runtime boundary that traces all agent actions and enforces policy as agents execute on Nvidia Vera CPUs (the company's first purpose-built CPU for agentic AI). As open-source software under Apache 2.0, OpenShell can be extended to third-party compute platforms including Arm and Intel.
The key architectural shift: instead of relying on the model or agent harness to self-limit, OpenShell creates an enforceable boundary outside the agent loop. This matters because every major agent incident to date has involved the agent working around its own guardrails at the application layer.
OpenShell is available now through Nvidia developer resources and GitHub.
Sentry: Out-of-Band Monitoring on BlueField-4
Sentry runs on Nvidia BlueField-4 DPUs (Data Processing Units) and provides continuous, in-silicon monitoring of agent behavior. Because it operates from an isolated, out-of-band trust domain, Sentry is invisible to both agents and attackers — it cannot be disabled or circumvented from within the agent's execution environment.
Sentry combines three capabilities:
- Threat detection — inspecting agent requests and responses in real time
- Hardware-based governance and enforcement — quarantining agents that attempt to move outside their boundaries in milliseconds
- Data access protection — enforcing granular, zero-trust access policies for data, tools, APIs, and services
Sentry is built on Nvidia DOCA (Data Center Infrastructure-on-a-Chip Architecture), which provides the programmable capabilities for attested telemetry, agent identity verification, and policy enforcement.
Ecosystem: 100+ Organizations, Including Anthropic, Microsoft, SpaceXAI
The platform launches with a broad coalition: Anthropic, Cisco, CrowdStrike, Dell, Figure, HPE, Hugging Face, JPMorganChase, Microsoft, Palantir, Palo Alto Networks, Perplexity, Red Hat, Salesforce, SAP, Scale AI, ServiceNow, SpaceXAI, and others.
Several integrations are already live or in progress:
- Anthropic: Claude Managed Agents runs the agent loop in a separate server from execution sandboxes. OpenShell and BlueField integrations add governance across those sandboxes.
- SpaceXAI: Using the platform for Cursor coding agents and Grok models.
- Scale AI: Incorporating the platform into Scale GenAI Portfolio for enterprise and government customers.
- Salesforce: Integrated OpenShell with Slack for agent activity visibility, audit events, and human approval workflows.
- SAP: Embedding OpenShell with Joule Studio runtime, contributing engineering work to OpenShell.
- Robotics leaders (Figure, Gecko Robotics, Skild AI): Embedding agent safety controls into autonomous physical systems.
- Infrastructure OS vendors (Canonical, SUSE, Red Hat): Integrating OpenShell and DOCA into widely used operating systems. Red Hat runs both on Red Hat AI Factory with Nvidia.
What This Means for MCP and Agent Infrastructure Builders
If you're building MCP servers, agent runtimes, or tooling that sits between models and external systems, this announcement signals three things:
1. The Application Layer Is Not a Security Boundary
The pattern across recent incidents is consistent: agents bypass application-layer controls. OpenShell and Sentry move enforcement to the runtime and hardware layers — places the agent cannot reach. For MCP server authors, this means your sandboxing and permission model must eventually integrate with runtime-level boundaries, not just in-process checks.
2. Agent Identity and Attestation Are Becoming Infrastructure Primitives
Sentry verifies agent identity and provides attested telemetry. This aligns with the direction MCP is heading: the 2026-07-28 spec made the protocol stateless, and agent identity/authorization is increasingly handled at the transport and infrastructure layer (OAuth, mTLS, SPIFFE). OpenShell's policy enforcement will likely expect verifiable agent identities — something MCP server authors should plan for.
3. Human-in-the-Loop Approval Is Becoming a Platform Feature
Salesforce's Slack integration shows the pattern: agent requests additional permissions, human approves or rejects, audit trail recorded. This is the operational model for production agents. If your MCP servers expose sensitive tools, you need an approval gateway — not just a boolean allow/deny config.
The Open Secure AI Alliance
Nvidia also highlighted the Open Secure AI Alliance, initiated alongside 120+ organizations and governed by the Linux Foundation. The alliance runs the Shared AI Findings Exchange (SAFE) — a coordinated vulnerability disclosure and best-practice sharing framework for agent security.
This is the closest thing the agent ecosystem has to a CVE/ISAC equivalent. If you operate MCP servers or agent infrastructure in production, monitoring SAFE advisories should be part of your security hygiene.
Availability and Next Steps
OpenShell and associated skills are available now via Nvidia's developer documentation and GitHub. Sentry requires BlueField-4 DPU hardware.
For developers building on MCP and Claude today, the immediate takeaway: the industry is converging on runtime-level enforcement as the security baseline. Application-layer guards (CLAUDE.md, pre-tool hooks, MCP server permissions) are necessary but not sufficient. The next wave of agent infrastructure will integrate with platforms like OpenShell — or implement equivalent boundaries themselves.
If this was useful, you can support my open-source work on Ko-fi or check out my services.