Lighthouse AdvisorySLED AI Adoption Intelligence
← Back to results

From the NVIDIA edition of September 28, 2026

Vendor claimEmergingNew this fortnight

Developer blog: Open Agent Safety Platform reference for continuous in-silicon agent monitoring

NVIDIA · AI agent security / infrastructure reference architecture · Global (vendor technical blog)

NVIDIA
Publisher
NVIDIA Developer Blog
Original publication
2026-09-28T00:00:00Z
Source retrieved
2026-09-29
Event date
2026-09-28
Read original source

What happened

Sep 28, 2026 NVIDIA Developer Blog details the Open Agent Safety Platform as a layered architecture: OpenShell sandboxed runtime with kernel-level isolation on Vera CPUs, plus Sentry on BlueField-4 DPUs using DOCA for out-of-band observability and real-time policy enforcement. It states five principles (verifiable policy, out-of-band enforcement, controlling the path to the model, scaling authority with reasoning visibility, shared responsibility) and notes that in Vera Rubin POD systems BlueField-4 sits on the node's path to the model for continuous monitoring.

Why it matters

Helps SLED architects map which controls are software-only versus DPU-path dependent before piloting campus/agency agents on NVIDIA pods.

Evidence and measured results

Opened developer blog dated Sep 28, 2026 by Myers/Watson/Golshan/Arkin; AI summary and body describe OpenShell+Sentry layering, DOCA, Vera Rubin POD BlueField-4 path, and shared-responsibility framing.

Limitations and uncertainty

Vendor architecture narrative; no independent campus measurements; POD-specific path claims may not transfer to heterogeneous clusters.

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
Buyers need to know which agent controls require BlueField-4 versus CPU-only OpenShell.
Stakeholders
  • infra architect
  • CISO
  • AI platform
Discovery questions
  • Are you on Vera Rubin POD?
  • Is BlueField-4 deployed?
  • Is host-compromise resilience a hard requirement?
Value hypothesis
Layered OpenShell+Sentry may match high-assurance agent designs if hardware path exists.
Potential engagement
Architecture workshop mapping POD vs mixed fleet control options.
Unsupported claims
  • Universal Sentry coverage on all GPU nodes
  • Measured campus outcomes in this blog

Pre-sales engineering

Role takeaway
Fit
Good fit when designing dual-layer controls on Vera/BlueField-4.
Architecture and integration
OpenShell on host CPU; Sentry/DOCA on BlueField-4 for out-of-band enforcement.
Prerequisites
  • Vera/BlueField-4 inventory
  • DOCA
  • policy/telemetry correlation
Constraints
  • POD path assumptions
  • vendor architecture without third-party test
Security
Out-of-band DPU path intended to survive host compromise—validate locally.
Proof of value
  • OpenShell deny works without DPU
  • Sentry alert/quarantine when DPU enabled

Delivery

Role takeaway
Work
Phase OpenShell then optional Sentry; write topology-specific runbooks.
Dependencies
  • DPU firmware/software
  • telemetry pipeline
  • agent catalog
Ownership
Infrastructure security architecture team.
Skills and adoption
Upskill on DOCA and sandbox operations.
Governance checkpoints
  • Threat model sign-off
  • Control-layer RACI
  • Change window for DPU enablement
Acceptance criteria
  • Layered test evidence
  • Topology diagram approved
Risks
  • Misapplied POD assumptions
  • Telemetry gaps

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?

Treat Sentry as path-to-model/DPU-layer control; OpenShell remains host-CPU sandbox even without DPU.

Governance

Who approves, reviews and stays accountable for outcomes?

Shared-responsibility model across labs, enterprises, and hardware providers must be translated into local RACI.

Security and privacy

What data, permissions and controls need testing?

Out-of-band DPU monitoring aims to remain effective if host is compromised—still vendor architecture claim pending local threat modeling.

Accessibility and workforce

Who is affected, and what skills or accommodations follow?

Architecture diagrams should distinguish CPU vs DPU controls for mixed-skill teams.

Procurement

What should contracts, pricing and exit terms secure?

BOM must separate OpenShell software from BlueField-4/DOCA dependencies.

Operating model

Which teams own the service once it runs?

Shared-responsibility language needs local ownership mapping before production agents.

Publication history

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

Stable resource ID: nvidia-devblog-agent-safety-sentry-20260928