Nvidia Launches Platform to Monitor and Constrain AI Agents
Summary
Nvidia introduced an open platform that combines software sandboxing and hardware monitoring to enforce limits on AI agents, including when a host system is compromised.
Key points
- Open Agent Safety Platform combines OpenShell, a policy-enforcing runtime, with Sentry, a hardware-based monitoring layer.
- OpenShell applies kernel-level controls to file and process activity, checks outbound requests against policy, and logs decisions.
- Sentry runs on Nvidia BlueField-4 DPUs and is designed to detect and block agent actions that exceed software-defined boundaries, even if the host is compromised.
- Agents may propose policy changes when enabled, but cannot approve their own requests; a policy prover checks permissions against operator-set limits.
- Nvidia said its tests found some agents with reduced safeguards spent up to two hours trying to obtain permission to modify a protected GitHub repository.
- Nvidia is working with Anthropic; SpaceXAI is using the technology, while Salesforce and SAP are integrating or planning to integrate OpenShell.
- Sentry protection can be enabled by software update on systems already running Nvidia Vera CPUs with BlueField-4; OpenShell is available through Nvidia developer resources and GitHub.
Article Details
- Event Type
- Launch of an AI agent safety and security platform
- Impact
- The platform is designed to monitor and constrain AI agents, enforce policy boundaries, and block unauthorized actions. Nvidia said Sentry can continue enforcing controls if an agent's host is compromised. SpaceXAI is already using the technology; other organizations are implementing it or plan to integrate it.
People
Vendors
Accentureembed OpenShell in its JouleStudio runtime. Other partners include CrowdStrike, Microsoft, Perplexity, Accenture, ServiceNow, and IBM to bring the technology to customers.AnthropicNvidia is working with Anthropic to implement the Open Agent Safety Platform, while SpaceXAI is already using the technology. Salesforce and Nvidia integrated OpenShell with Slack to allow teams to view agent activity,CrowdStrikepermissions. SAP is planning to embed OpenShell in its JouleStudio runtime. Other partners include CrowdStrike, Microsoft, Perplexity, Accenture, ServiceNow, and IBM to bring the technology to customers.IBMJouleStudio runtime. Other partners include CrowdStrike, Microsoft, Perplexity, Accenture, ServiceNow, and IBM to bring the technology to customers.MicrosoftSAP is planning to embed OpenShell in its JouleStudio runtime. Other partners include CrowdStrike, Microsoft, Perplexity, Accenture, ServiceNow, and IBM to bring the technology to customers.NvidiaNvidia has announced a policing system that will watch and rein in artificial intelligence (AI) agents before they go rogue. The Open Agent Safety Platform combines open source software and a reference system design toPerplexityplanning to embed OpenShell in its JouleStudio runtime. Other partners include CrowdStrike, Microsoft, Perplexity, Accenture, ServiceNow, and IBM to bring the technology to customers.SalesforceAnthropic to implement the Open Agent Safety Platform, while SpaceXAI is already using the technology. Salesforce and Nvidia integrated OpenShell with Slack to allow teams to view agent activity, audit events, andSAPteams to view agent activity, audit events, and approve or reject agent requests for more permissions. SAP is planning to embed OpenShell in its JouleStudio runtime. Other partners include CrowdStrike, Microsoft,ServiceNowin its JouleStudio runtime. Other partners include CrowdStrike, Microsoft, Perplexity, Accenture, ServiceNow, and IBM to bring the technology to customers.SpaceXAINvidia is working with Anthropic to implement the Open Agent Safety Platform, while SpaceXAI is already using the technology. Salesforce and Nvidia integrated OpenShell with Slack to allow teams to view agent activity,
Products
BlueField-4on some x86 and Arm CPUs. The second is Sentry, a hardware-based watchdog layer that runs on Nvidia's BlueField-4 data processing units (DPUs) and monitors agents' actions.DOCAagent's host, Sentry can still block agents even when the host is compromised. Sentry is built on Nvidia's DOCA software, which inspects agent requests and responses, provides attested telemetry, verifies agentJouleStudioand approve or reject agent requests for more permissions. SAP is planning to embed OpenShell in its JouleStudio runtime. Other partners include CrowdStrike, Microsoft, Perplexity, Accenture, ServiceNow, and IBM toOpen Agent Safety Platformthat will watch and rein in artificial intelligence (AI) agents before they go rogue. The Open Agent Safety Platform combines open source software and a reference system design to keep AI agents within setOpenShellsit between enterprise IT systems and AI agents to define, monitor, and contain agents. One layer is OpenShell, an open source runtime that Nvidia originally introduced in March to sandbox agents and enforce policySentryCodex, Claudo Code, Pi, Hermes, and other agents and runs on some x86 and Arm CPUs. The second is Sentry, a hardware-based watchdog layer that runs on Nvidia's BlueField-4 data processing units (DPUs) andSlackPlatform, while SpaceXAI is already using the technology. Salesforce and Nvidia integrated OpenShell with Slack to allow teams to view agent activity, audit events, and approve or reject agent requests for moreVeraHardware enforces the agentic policing system. Nvidia's Arm-based Vera CPU runs OpenShell, which sets boundaries in which agents can run safely. The BlueField-4 data processing unit, which also includes an Arm CPU,