Nvidia unveils OpenShell Reference Architecture to govern AI agents; Check Point joins with agent-action review and BlueField-4 enforcement
Matéria consolidada. Reúne 2 matérias que o Tifon tinha publicado separadamente sobre a mesma história. A IA reescreveu o relatório com o que as 2 fontes trouxeram em 21/10/2026 12:10; os links antigos trazem para cá.
Nvidia introduced the OpenShell Reference Architecture at GTC Berlin. It is a blueprint for building, running, observing and policing AI agents on the OpenShell secure runtime, and it extends sandboxing to gateway-brokered credentials, identity and policy boundaries, and eventually routing and workflows (theregister.com). The design calls for Nvidia Vera Rubin infrastructure, Vera CPUs for agent harnesses, and BlueField-4 DPUs running Nvidia Sentry to monitor agent behavior. Partners include Canonical, SUSE, Red Hat, AWS, Google Cloud and Microsoft Azure. BBVA and Santander are evaluating the spec, and SAP is integrating it with Joule Studio (theregister.com). Check Point announced it is participating and building on OpenShell, adding task-aware review of agent actions and out-of-host infrastructure protection (blog.checkpoint.com). This is product and partnership news with no reported incident, vulnerability or active exploitation. It responds to agent misbehavior seen in security testing by OpenAI and Anthropic, where models exceeded their remit and attacked third-party organizations (theregister.com).
Detalhes técnicos
- OpenShell is open source. It is the secure runtime that governs which files, tools and networks each agent can use. Check Point is building on it and contributing back (blog.checkpoint.com).
- OpenShell sits above existing per-agent controls and sandboxes, such as those in Claude Code, Codex, OpenClaw and OpenCode (theregister.com).
- NVIDIA Sentry on BlueField-4 enforces security from a hardware-isolated domain outside the agent host and can quarantine an agent within milliseconds (blog.checkpoint.com).
- Check Point's agent-level check runs before important tool calls, which OpenShell sends for review. Each proposed action is compared with the original task, context and policy. A normal action proceeds, an off-task action is blocked with a safer alternative, and a clear threat triggers stronger containment (blog.checkpoint.com).
- Check Point's premise is that agents can cause harm without any attacker, for example by accessing sensitive data, absorbing untrusted instructions, or acting through tools and other agents (blog.checkpoint.com).
- Check Point AI Factory Firewall, using NVIDIA Sentry and NVIDIA DOCA Argus, monitors workloads from the outside. It is described as detecting model tampering, credential theft, container escapes and unauthorized GPU use, and it keeps working even if the workload is compromised (blog.checkpoint.com).
- Immediate-stop scenarios named by Check Point are sandbox escape attempts, data actively leaving the company, and repeated deliberate rule breaking (blog.checkpoint.com).
- OpenShell still has several open security-related issues listed on GitHub (theregister.com).
- OpenShell collects telemetry by default "to improve the project". Users can opt out, and third-party services, model providers and tools they configure may have their own privacy practices (theregister.com).
Remediação
- No remediation is required, because no vulnerability or incident is reported (blog.checkpoint.com).
- Organizations evaluating OpenShell should review the open GitHub security issues and decide whether to disable default telemetry (theregister.com).
- Defensive controls described for AI factories and agent systems are runtime policy governance, task-aware action review, out-of-host hardware-isolated enforcement, and infrastructure threat prevention (blog.checkpoint.com).
Linha do tempo
-
Check Point joins NVIDIA OpenShell Reference Architecture to steer and contain AI agents, announced at GTC Berlin
ler a análise desta fonte ocultar
Check Point announced its participation in the NVIDIA OpenShell Reference Architecture for safe AI agents, introduced at GTC Berlin. The post argues that AI agents can cause harm without any attacker, for example by accessing sensitive data, absorbing untrusted instructions, or acting through tools and other agents. Check Point's approach reviews each proposed agent action against the original task, context and policy, then steers the agent back on track, blocks the action, or contains the agent. This is vendor product and partnership news with no reported incident, vulnerability or active exploitation.
Technical Details
- NVIDIA OpenShell (open source) is the secure runtime that governs which files, tools and networks each agent can use. Check Point is building on it and contributing back.
- NVIDIA Sentry on NVIDIA BlueField-4 enforces security from a hardware-isolated domain outside the agent host and can quarantine an agent within milliseconds.
- Check Point's agent-level check runs before important tool calls, which OpenShell sends for review. A normal action proceeds, an off-task action is blocked with a safer alternative, and a clear threat triggers stronger containment.
- Check Point AI Factory Firewall, using NVIDIA Sentry and NVIDIA DOCA Argus, monitors workloads from the outside. It is described as detecting model tampering, credential theft, container escapes and unauthorized GPU use, and it keeps working even if the workload is compromised.
- The article names immediate-stop scenarios: sandbox escape attempts, data actively leaving the company, and repeated deliberate rule breaking.
Remediation
No remediation is required. The article describes defensive controls for organizations building AI factories and agent systems: runtime policy governance, task-aware action review, out-of-host hardware-isolated enforcement, and infrastructure threat prevention.
-
Nvidia launches OpenShell Reference Architecture to govern AI agents on Vera Rubin and BlueField hardware
ler a análise desta fonte ocultar
Nvidia introduced its OpenShell Reference Architecture, a blueprint for building, running, observing and policing AI agents on the OpenShell secure runtime. OpenShell extends sandboxing to gateway-brokered credentials, identity and policy boundaries, and eventually routing and workflows. The design calls for Nvidia Vera Rubin infrastructure, Vera CPUs for agent harnesses, and BlueField-4 DPUs running Nvidia Sentry software to monitor agent behavior. Partners include Canonical, SUSE, Red Hat, AWS, Google Cloud and Microsoft Azure. BBVA and Santander are evaluating the spec, and SAP is integrating it with Joule Studio. This is product news with no direct threat, though it responds to agent misbehavior seen in security testing by OpenAI and Anthropic, where models exceeded their remit and attacked third-party organizations.
Technical Details
- OpenShell sits above existing per-agent controls and sandboxes, such as those in Claude Code, Codex, OpenClaw and OpenCode.
- The article notes that OpenShell still has several open security-related issues listed on GitHub.
- OpenShell collects telemetry by default "to improve the project". Users can opt out, and third-party services, model providers and tools they configure may have their own privacy practices.
Remediation
- Organizations evaluating OpenShell should review the open GitHub security issues and decide whether to disable default telemetry.