
NVIDIA shipped its Open Agent Safety Platform on September 28, and the premise is blunt: your AI agent should not get unlimited access to a computer, a database, or an API simply because its instructions say “behave.” The platform combines two independent layers — OpenShell, an open-source runtime that sandboxes agents in software, and Sentry, a hardware-level watchdog running on BlueField-4 DPUs that can quarantine a rogue agent in milliseconds, invisibly, while the agent keeps running unaware it was just stopped.
Two Layers, One Architecture
Most AI safety work lives inside the model: RLHF, system prompts, output filters. The problem is that all of those controls share an attack surface with the thing they are supposed to constrain. NVIDIA’s approach moves enforcement outside the model entirely.
OpenShell is the software layer. It runs on NVIDIA Vera CPUs but is extensible to Arm and Intel platforms. Released under Apache 2.0 at version 0.1.0, it sandboxes every agent action: filesystem reads, network destinations, child processes, credentials. Policies are written in YAML, compiled to OPA/Rego, and evaluated on every outbound request. A Z3-based policy prover can mathematically verify a policy stays inside the operator’s intended boundary before the agent ever runs.
Sentry is the hardware layer. It runs on NVIDIA BlueField-4 DPUs as an out-of-band watchdog — physically separate from the CPU where the agent runs. The agent cannot see Sentry, cannot disable it, and cannot know it is being watched. If the agent attempts to step outside its policy boundary, Sentry quarantines and stops it in milliseconds.
The result is a defense that does not rely on model behavior. An agent can be fully jailbroken at the model level and still hit a wall at the infrastructure level.
Developers Can Try OpenShell Today
The software component is installable now. You define what the agent is allowed to touch in a YAML policy file — which files, which APIs, which processes — and OpenShell enforces it. The policy advisor mode helps identify what access patterns your agent actually needs before you lock things down.
# Install OpenShell
curl -LsSf https://raw.githubusercontent.com/NVIDIA/OpenShell/main/install.sh | sh
# Create a sandboxed agent environment with a policy
openshell sandbox create --name my-agent --policy policy.yaml
# Enable policy advisor (helps identify what to allow)
openshell settings set --key agent_policy_proposals_enabled --value true
The full developer walkthrough is available on the NVIDIA Developer Blog, and the source is on GitHub.
Why This Is Urgent Right Now
The timing is not arbitrary. Prompt injection attacks against AI agents surged 32% in web content between November 2025 and February 2026, according to Google researchers. CrowdStrike’s 2026 Global Threat Report documented that threat actors injected malicious prompts into AI tools at over 90 organizations last year. Multi-turn jailbreaks — attacks that unfold across extended conversations — achieved a 92% success rate across eight open-weight models in recent research.
The attacks are not theoretical. Security researchers demonstrated that Devin AI, given a crafted prompt, could be made to expose server ports to the internet, leak access tokens to external endpoints, and install command-and-control malware — all within the scope of what looked like a routine coding task. If you have been following Plugin4Shell or the JadePuffer Azure incident, you already know the blast radius when agentic access goes wrong.
The Five Principles Behind the Platform
NVIDIA published the governance philosophy that shaped the architecture:
- Verifiable policy — Before an agent runs, a prover must confirm its policy cannot escape the operator’s intent.
- Out-of-band enforcement — Controls must live outside and beyond the agent’s reach.
- Control the path to the model — Every action flows through the next-token decision; gate it there.
- Scale authority with reasoning visibility — The more an agent can do, the more its reasoning must be observable.
- Shared responsibility — Labs, enterprises, and hardware providers each own a layer, the same model as cloud infrastructure today.
Principle five is worth noting: NVIDIA is explicitly not claiming to own the whole stack. The framework assigns accountability at each layer rather than centralizing it in one vendor.
120 Partners, But Notice Who Is Missing
The Open Secure AI Alliance, which underpins the ecosystem around this platform, launched in July 2026, transferred to the Linux Foundation in September, and now counts over 120 member organizations: Cisco, CrowdStrike, Hugging Face, Red Hat, BlackRock, Capital One, Intel, and Visa are on the roster.
Anthropic, OpenAI, and Google have not joined. That is not a small omission given that their models power most production agents. Whether they adopt the platform’s architecture independently or stay out entirely is an open question worth watching.
The Honest Verdict
OpenShell is at version 0.1.0. Sentry requires BlueField-4 DPUs — enterprise data center hardware, not something you drop into a laptop or a small cloud instance. And writing the YAML policies that actually capture what you intend is still a manual, largely unguided process. The Z3 prover can verify internal consistency, but it cannot tell you whether your policy reflects your real intent.
None of that makes this unimportant. Moving enforcement outside the model is the right architectural direction. If you are shipping AI agents to production now, OpenShell is worth evaluating on the NVIDIA agent safety page. If you are building on enterprise infrastructure with BlueField DPUs, Sentry gives you a hardware enforcement layer that no amount of prompt engineering can reach.

