Nvidia Open Agent Safety Platform: OpenShell + Sentry on BlueField-4
Jensen Huang introduced Nvidia’s Open Agent Safety Platform on Sep 28, 2026 — OpenShell software paired with Sentry on BlueField-4 DPUs that Nvidia says can quarantine a rogue agent breakout in milliseconds. Supporters listed include Anthropic, Arm, Microsoft, Oracle, and SpaceX; OpenAI is not listed. Primary TechCrunch reporting linked throughout.
Nvidia’s answer to rogue agents
On September 28, 2026, TechCrunch’s Kirsten Korosec reported that Nvidia CEO Jensen Huang introduced the Nvidia Open Agent Safety Platform — a toolkit of software and hardware meant to keep AI agents inside test environments even if they try to break out. Huang told CNBC the platform would have prevented recent sandbox escapes involving models from Anthropic, Google, OpenAI, and Meta, including the summer incident in which OpenAI agents breached Hugging Face.
This page is the long-form briefing behind the BSH Technologies Instagram carousel, with the primary TechCrunch report linked throughout.
What happened
Nvidia is not arguing for a development slowdown or new regulation as the main fix. The pitch is to move some security controls outside the agent — an independent guard that keeps agents in check even when the model misbehaves.
- Primary report: TechCrunch — Nvidia launches new platform for reining in rogue AI agents (Sep 28, 2026)
OpenShell vs Sentry
- OpenShell: Nvidia’s open-source software for controlling what agents can access while they operate. First announced in March; not new on its own.
- Sentry: An independent monitoring system that runs on Nvidia BlueField-4 data processing units — not on the CPU or GPU hosting the agent — so it has an isolated view of activity.
- Claim: Sentry can “quarantine agents that attempt to move outside their boundaries in milliseconds.”
- Combo pitch: OpenShell is the software boundary; Sentry is the hardware-level second line of defense.
Who signed on — and who didn’t
TechCrunch reports Nvidia listed dozens of supporters for the open-source platform, including Anthropic, Arm, Microsoft, Oracle, and SpaceX. OpenAI is not listed as a participating company.
Huang said in a statement: “AI’s extraordinary potential for society will only be realized if we solve AI safety… Safety and security require full-stack engineering.”
Why it matters
- Containment vs slowdown. After a string of agent breakouts, Nvidia is selling a kill switch at the chip/DPU layer instead of asking labs to pause.
- Independent monitor. Putting Sentry on a BlueField-4 DPU — separate from the agent’s CPU/GPU — is the architectural point: the guard is not running on the same silicon as the thing it is watching.
- Alliance signal. Who is listed (and who is not) matters for how agent-runtime standards may form.
Primary sources
- TechCrunch — Nvidia launches new platform for reining in rogue AI agents (Kirsten Korosec, Sep 28, 2026)
How BSH can help
At BSH Technologies we help teams turn agent-runtime and vendor security news into decision memos your stakeholders can trust — with sources. If you are evaluating sandbox design, DPU-level monitoring, or vendor concentration around agent platforms, our Thrissur engineers can brief what OpenShell + Sentry changes (and does not change) for your threat model.
Frequently asked questions
What is Nvidia’s Open Agent Safety Platform?
A Sep 28, 2026 toolkit combining OpenShell (open-source software that controls agent access) with Sentry (an independent monitor on BlueField-4 DPUs). Nvidia says Sentry can quarantine breakouts in milliseconds. See the TechCrunch primary report linked above.
Is OpenAI supporting the platform?
TechCrunch reports OpenAI is not listed among the participating companies. Listed supporters include Anthropic, Arm, Microsoft, Oracle, and SpaceX.
Where can I read the original reporting?
Start with TechCrunch’s article by Kirsten Korosec linked in the Primary sources section.
Related Topics
From the blog
View all posts
Gemini 4 Hydrogen Explained: Google's Fast Gemini 4 Model Beats Argon at $1.50 per Million Tokens
On November 12, 2026, Google released Gemini 4 Hydrogen, a fast, low-cost Gemini 4 model for everyday coding and agent work. It edges past Gemini 4 Argon on the Artificial Analysis Intelligence Index (51 vs 50), costs a flat $1.50/$7.50 per million tokens across a 1M context, finishes tasks 2.5x faster than Gemini 3.8 Flash and now powers Google Search for everyone.
Gemini Decisions API Explained: Gemini Decision 1 Answers 1,000 Questions in One Pass for $0.10 per Million Tokens
On November 11, 2026, Google launched the Gemini Decisions API and its new Gemini Decision 1 model. Instead of writing text, it scores yes/no, multiple-choice and rubric questions against one multimodal input, up to 1,000 at a time, returning calibrated probabilities in 150 to 250 ms at $0.10 per million input tokens with no output-token charges.