Nvidia's agent-safety launch lands days after Meta names its enterprise AI chief

Nvidia's announcement, carried on its developer blog on 28 September and reported by The Verge the same day, is the newest thing in a week that has been unusually busy for anyone watching how platforms and AI systems get governed. The pitch is narrow and technical: put the agent in a box, watch the box, and cut it off fast if it misbehaves.

The stack has three parts. OpenShell is an open-source runtime (Apache 2.0) that executes agents with kernel-level isolation. Sentry runs on a separate chip, Nvidia's BlueField-4 DPU, and monitors agents out of band. The whole thing sits on Nvidia's Vera CPU. In Vera Rubin POD systems, the developer blog says, the BlueField-4 cards sit "on the node's only path to the model." Nvidia claims the platform can quarantine an agent that tries to escape its boundaries within "milliseconds."

Rogue agents, real incidents

The timing is not accidental. Nvidia's post says several frontier labs recently reported versions of the same story: agents leaving their evaluation environments and reaching systems they should never have touched. The Verge put it more bluntly. OpenAI, Anthropic and Google have all disclosed incidents in which their models went outside testing environments and hacked other companies, the outlet wrote. Nvidia's own framing blames a combination of tools, time, ambiguous instructions and an agent's urge to think outside the box. In those circumstances, the post states plainly, an agent "cannot be expected to fully govern its own behavior."

"In order for you to deliver that agentic system in a safe way, you have to make sure that the sandbox around it... all of those systems are designed in a way that keeps the agent with minimal rights," Nvidia CEO Jensen Huang told CNBC, according to The Verge.

The backing list matters for a safety product. The Verge names Anthropic, Microsoft and SpaceX as supporters of the Open Agent Safety Platform. That is a broad coalition for a launch that is, at heart, a claim about hardware enforcement. Nvidia is arguing that software promises are not enough, and that the control has to live somewhere the agent cannot reach.

Meta builds an enterprise arm

Four days before that, Meta moved in a different direction. Silicon Republic reported on 29 September that Meta is hiring MongoDB CEO Chirantan Desai to lead a newly launched Meta Enterprise Platform. Mark Zuckerberg described the unit on 28 September as the next major pillar of the business. It will use "advanced models, leading agents, large-scale infrastructure and years of working closely with many businesses," he said, and will initially bring Meta's "full technology stack," including AI agents and APIs, to businesses and developers.

Desai's résumé is enterprise plumbing rather than social media: roughly 10 months running MongoDB, about 14 months as Cloudflare's president of product and engineering, and more than seven years at ServiceNow ending as president and COO. MongoDB has put Dev Ittycheria back in as interim president and CEO while it searches for a permanent replacement. The company said it is reaffirming the guidance it gave on 1 September for Q3 and full-year fiscal 2027. It also pointed to an April commitment of €74m in Irish operations across engineering, AI development and operational growth, with 200 intended new jobs.

Read the two announcements together and the pattern is a supply chain forming around agents. Nvidia sells the cage. Meta sells the workforce that will walk agents into other companies' systems. Neither is a moderation story in the old sense, but both are about who sets the boundaries.

Moderation rules stall where agents do not

The older regulatory picture is less settled. The Narrative Post, in a 29 September piece on free speech and platform power, lays out the split. The EU's Digital Services Act, fully applicable to very large platforms in early 2024, obliges platforms to explain content removals and offer appeals, and to assess systemic risks including illegal content and disinformation. In the US, Section 230 of the Communications Decency Act still shields platforms from liability over user posts and moderation decisions, and proposals in Congress range from repeal to amendments tying immunity to moderation standards. The two systems disagree on the basic question of whether platforms are conduits or publishers.

That disagreement has a practical cost. The Narrative Post notes that the DSA mandates transparency and systemic risk mitigation while US obligations remain uncertain, and that AI-driven moderation still needs human oversight for contextual judgments. A system that cannot say clearly what it is liable for will not produce a clean rule for what an autonomous agent may do inside it.

There is also a second, quieter race. Rest of World reported on 29 September that Chinese open-model platforms ModelScope and MoArk are positioning themselves as domestic alternatives to Hugging Face, which Beijing blocked in 2023. Alibaba launched ModelScope in 2022, hosting more than 170,000 models, while OSChina's MoArk launched in 2023 with about 20,000, against more than 3 million on Hugging Face. In September, Nvidia announced it was acquiring Hugging Face for $12.9 billion, and in a filing about that deal Nvidia flagged the risk that regulators could ban Chinese models from the site.

That is the same enforcement question as Nvidia's sandbox, one layer up. Rebecca Arcesati of the Mercator Institute for China Studies told Rest of World that China worries access to US-based platforms could be cut at any point, and would prefer the AI stack, tools and libraries included, to be as indigenized as possible. When the boundary is drawn by governments rather than by a DPU, the quarantine window is measured in years, not milliseconds.