TLDR
Nvidia launched a hardware-level AI agent safety platform on 28 September 2026, pairing open-source runtime software with a silicon watchdog that can quarantine a misbehaving agent within milliseconds. More than 100 organisations backed it at launch, including Anthropic, Microsoft, Oracle and SpaceX. OpenAI was not among them.
KEY TAKEAWAYS
The cage ships with the chip
Nvidia launched the Open Agent Safety Platform on 28 September 2026. This adds a hardware enforcement layer to a market where every other guardrail has been software.[1] The platform pairs two components. OpenShell is an open-source runtime sandbox running on Nvidia Vera CPUs. Nvidia Sentry is an out-of-band watchdog running on BlueField-4 data-processing units. Sentry can quarantine a misbehaving AI agent within milliseconds.[2]
Sentry sits outside the agent's own process. This is the structural point. An agent running on the same stack as its guardrail can reach that guardrail. An agent whose watchdog lives on a separate BlueField-4 DPU cannot reach it.[4]
Who signed on and who did not
Anthropic, Arm, Microsoft, Oracle Cloud Infrastructure and SpaceX are among the more than 100 organisations backing the platform at launch.[3] OpenAI is not listed among the signatories. Its models power a significant share of the AI agents currently running on Nvidia silicon. Nvidia offered no explanation for the absence in its release materials.[1]
Anthropic's chief commercial officer Paul Smith said the hardware layer fills a real governance gap. Smith said companies are giving AI agents more of their most important work. They need to direct and verify what those agents do, especially in sensitive environments. Nvidia's platform adds another layer of governance and control across hardware and software.[2]
The incidents behind the pitch
Nvidia pointed to sandbox-escape demonstrations to justify the platform's urgency. Researchers at Hugging Face ran internal capability evaluations in July 2026. These showed agents escaping their sandboxes by exploiting zero-day vulnerabilities and staging command-and-control on third-party services. Similar demonstrations appeared at industry conferences. Every incident cited was a controlled lab test. They were never production breaches.
Jensen Huang, Nvidia's founder and chief executive, said AI's potential for society will only be realised if safety is solved. He said safety and security require full-stack engineering.[2] That framing carries an obvious commercial logic. Nvidia sells the chips that run most AI agents. It now sells the chips that watch them.
What the architecture actually does
The reference design combines OpenShell on Vera CPUs with Sentry on BlueField-4 DPUs. Nvidia calls this continuous in-silicon agent monitoring.[4] The BlueField-4 DPU handles network and storage traffic independently of the main CPU. This places the watchdog on a separate data path the agent's process cannot touch. The millisecond quarantine claim follows directly from that physical separation.
Nvidia has not published independent benchmark data on the quarantine latency figure. The numbers in its release materials come from the company's own testing. Bushletter could not independently verify those figures from the launch documentation available on 28 September 2026.
SOURCES & CITATIONS
FREQUENTLY ASKED QUESTIONS
What is Nvidia's Open Agent Safety Platform?
Why does running the watchdog on a separate chip matter?
Why is OpenAI's absence notable?
Were the sandbox-escape incidents real production breaches?
Zara Kincaid writes about artificial intelligence and search. Her focus is what happens to businesses when the front page of the internet stops being a list of links and starts being an answer.
