Lighthouse AdvisorySLED AI Adoption Intelligence
← Back to results

From the NVIDIA edition of September 28, 2026

Vendor claimEmergingNew this fortnight

NVIDIA launches Open Agent Safety Platform with OpenShell runtime and BlueField-4 Sentry

NVIDIA · AI agent security / enterprise platform software · Global (vendor press; public-sector cooperation framed)

NVIDIA
Publisher
NVIDIA Investor Relations / GlobeNewswire
Original publication
2026-09-28T00:00:00Z
Source retrieved
2026-09-29
Event date
2026-09-28
Read original source

What happened

On Sep 28, 2026 NVIDIA announced the Open Agent Safety Platform: open-source OpenShell secure runtime (broadly available) that enforces agent policy boundaries on CPUs including NVIDIA Vera, plus a Sentry reference design—an out-of-band watchdog on BlueField-4 DPUs that NVIDIA says can quarantine agents leaving their software boundary in milliseconds. The release cites recent incidents where agents circumvented application-layer controls and frames cooperation with industry, researchers, and public-sector organizations; partner names include Anthropic, Cisco, CrowdStrike, Dell, HPE, Microsoft, Palantir, Red Hat, Salesforce, SAP, Scale AI, ServiceNow, and SpaceXAI.

Why it matters

Campus and agency teams piloting long-running coding/research agents need enforceable runtime and optional DPU controls beyond app-layer allowlists; this is a vendor platform announcement, not a measured SLED deployment outcome.

Evidence and measured results

Opened investor press release dated September 28, 2026 (GlobeNewswire Santa Clara). News summary and body describe OpenShell, Sentry/BlueField-4, milliseconds quarantine claim, Vera CPU pairing, Arm/Intel extensibility, and public-sector framing.

Limitations and uncertainty

Vendor announcement without independent SLED measured outcomes; quarantine latency and partner depth are company-stated; OpenShell maturity (0.1.0 per companion coverage) implies early software.

Cross-source patterns this evidence supports

Put this evidence to work

Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-29; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.

Sales

Role takeaway
Customer problem
Long-running agents can bypass application-layer controls and need enforceable boundaries.
Stakeholders
  • CISO
  • AI platform lead
  • research computing
  • procurement
Discovery questions
  • Which agent frameworks are in pilot?
  • Do you have BlueField-4/Vera?
  • Who owns agent policy approval?
Value hypothesis
OpenShell plus optional Sentry may reduce agent escape risk if policies and hardware match the reference design.
Potential engagement
Time-boxed OpenShell policy PoC on one non-prod agent class with quarantine/deny drills.
Unsupported claims
  • FedRAMP ATO from this announcement
  • Guaranteed millisecond containment in every environment

Pre-sales engineering

Role takeaway
Fit
Fit for agent fleets needing runtime sandboxing; Sentry only where BlueField-4 path exists.
Architecture and integration
Deploy OpenShell gateway/sandbox/supervisor; optionally add Sentry on BlueField-4 as out-of-band monitor.
Prerequisites
  • OpenShell 0.1.x access
  • agent identity
  • policy store
  • BlueField-4 if Sentry
Constraints
  • Early software maturity
  • Sentry is reference design
  • no ATO claimed
Security
Enforcement outside agent workload; credentials should not be visible inside the sandbox.
Proof of value
  • Deny unauthorized API write
  • Quarantine drill logged
  • Policy change cannot self-approve

Delivery

Role takeaway
Work
Pilot OpenShell policies, logging, and optional Sentry enablement on approved hardware.
Dependencies
  • Dev resources/GitHub
  • CPU/DPU inventory
  • IdP
  • SIEM sink
Ownership
Platform security with AI platform engineering.
Skills and adoption
Train operators on policy authoring and quarantine runbooks.
Governance checkpoints
  • Policy ownership RACI
  • Change control for policy prover exceptions
  • Data-handling review
Acceptance criteria
  • Documented deny/quarantine tests
  • Audit events retained
  • Rollback path defined
Risks
  • Over-scoping production credentials
  • Assuming DPU coverage without inventory

Implementation considerations

Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.

Architecture and integration

What must connect, and where does the AI sit in the workflow?

Agent stacks gain a CPU-side sandbox/runtime boundary (OpenShell) and an optional DPU out-of-band enforcement path (Sentry on BlueField-4); not a model-weight change.

Governance

Who approves, reviews and stays accountable for outcomes?

Who defines agent policies, who approves policy-change proposals, and whether DPU/Sentry is mandatory for higher-risk agent classes become new governance questions.

Security and privacy

What data, permissions and controls need testing?

Moves enforcement outside the agent/app layer; does not by itself create FedRAMP ATO or replace system-boundary authorization.

Accessibility and workforce

Who is affected, and what skills or accommodations follow?

Policy and quarantine runbooks need plain-language paths for non-security developers running agents.

Procurement

What should contracts, pricing and exit terms secure?

Treat as software/reference-design procurement plus optional DPU dependency—not a certified government product claim.

Operating model

Which teams own the service once it runs?

Establish joint AI-platform and security ownership for agent policy lifecycle.

Publication history

  1. 2026-09-28NVIDIA · Issue 237 resources
Read preserved resource versions (JSON)

Stable resource ID: nvidia-open-agent-safety-pr-20260928