AI NexusAI Nexus
Product update

NVIDIA launches an open agent safety platform: OpenShell software boundaries plus a BlueField-4 hardware watchdog

Event time 3 independent sourcesEditorial score 60/100Updated here

The short version

NVIDIA announced the Open Agent Safety Platform, made up of the open-source OpenShell runtime and the Sentry reference system design, which constrains AI agents with software boundaries and hardware monitoring. NVIDIA says Sentry, running on BlueField-4 DPUs, can isolate an agent that crosses its boundaries within milliseconds; the platform software, including OpenShell and skills, is available from NVIDIA's developer resources page and GitHub.

What changed

NVIDIA combined the open-source OpenShell runtime with the new Sentry reference system design into one agent safety platform: OpenShell sets and enforces boundaries on the CPU, while Sentry independently monitors agent behavior out of band on a BlueField-4 DPU.

What it means for you

Teams deploying coding or business agents in enterprise environments can evaluate OpenShell as an enforceable permission boundary outside the model and the agent framework. Sentry is a reference system design that depends on BlueField-4 DPUs; what can be downloaded and used today is software such as OpenShell.

Participants listed in the press release include Anthropic, Cisco, CrowdStrike, Hugging Face, Microsoft, Red Hat, Salesforce and SAP. Among them, Salesforce and NVIDIA have connected OpenShell to Slack, where teams can view agent activity and audit events and approve or deny an agent's requests for extra permissions.

Note: The end of the press release states that many of the products and features described are at various stages and are provided "if and when available"; Sentry is a reference system design, not an out-of-the-box product.

Fact check

  • VerifiedThe platform consists of NVIDIA OpenShell open-source software and the NVIDIA Sentry reference system designEvidence
  • VerifiedOpenShell sets runtime boundaries for agents, tracks all actions and enforces policyEvidence
  • VerifiedSentry runs on BlueField-4 DPUs, monitors agents continuously out of band, and can isolate an agent that tries to cross its boundaries within milliseconds"Milliseconds" is NVIDIA's own wording; there is no independent testing yet.Evidence
  • VerifiedOpenShell can extend to third-party compute platforms such as Arm and IntelEvidence
  • VerifiedThe platform software, including OpenShell and skills, is available through NVIDIA's developer resources page and GitHubEvidence
  • VerifiedAnthropic is working with NVIDIA, and Claude Managed Agents is integrated with OpenShell and BlueFieldEvidence
  • VerifiedParticipants listed in the press release include Anthropic, Cisco, CrowdStrike, Hugging Face, Microsoft, Red Hat, Salesforce and SAP; Salesforce and NVIDIA have connected OpenShell to Slack, where agent activity can be viewed and permission requests approvedEvidence
  • VerifiedThe press release states that many of the products and features described are at various stages and are provided "if and when available"Evidence
  • VerifiedNVIDIA's press release shows only the date 2026-09-28; neither the page nor its metadata gives a specific publication timeThis site treats the event as date-only per the source and does not fill in a specific time.Evidence

Coverage timeline

  1. CoverageThe Verge
    Nvidia says its new AI safety platform can contain rogue agents within ‘milliseconds’
  2. CoverageTechCrunch
    Nvidia launches new platform for reining in rogue AI agents
  3. Primary sourceNVIDIA
    NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment