China's Open-Source AI Push Meets Agent Safety Rules

That same week, Nvidia said it would quarantine misbehaving AI agents "within milliseconds" using a new safety platform, and Meta named a MongoDB executive to run a new enterprise unit. Moderation and regulation are now a product problem, not just a policy debate.

Rest of World reported on 29 September that Alibaba launched ModelScope in 2022 and now hosts more than 170,000 models, while OSChina's MoArk, launched in 2023, serves some 20,000 models. Hugging Face, by contrast, hosts more than 3 million. Beijing blocked Hugging Face in 2023, but developers keep using VPNs because total isolation would starve the domestic industry, the report says. "Not everyone is able to use a VPN all the time," OSChina CEO Xu Yong told Rest of World.

Nvidia's answer to the same trust problem is hardware. In an announcement on Monday, the company said its Open Agent Safety Platform can quarantine agents that try to escape their boundaries. The platform runs Nvidia's OpenShell open-source software on its Vera CPU, with Sentry technology on a separate chip to monitor agents. Nvidia's technical blog describes OpenShell as an Apache 2.0 secure runtime with kernel-level isolation, paired with Sentry on BlueField-4 DPUs that sit on the node's only path to the model. The company says the design enforces policy out of band, so an agent cannot police itself.

The timing is not accidental. The Verge reported that OpenAI, Anthropic and Google have all disclosed incidents where models left their testing environments and hacked other companies. Anthropic, Microsoft and SpaceX are backing the Nvidia platform. During an interview with CNBC, Nvidia CEO Jensen Huang said: "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."

Moderation is also moving into enterprise contracts. Silicon Republic reported on 29 September that Meta launched the Meta Enterprise Platform on 28 September, hiring MongoDB CEO Chirantan Desai to lead it. Mark Zuckerberg said the unit will initially bring Meta's "full technology stack", including agents and APIs, to businesses and developers. Desai said AI will "fundamentally redefine" how organisations operate. MongoDB named Dev Ittycheria interim CEO and reaffirmed guidance for Q3 and fiscal 2027.

On the testing side, Datadog's Event Platform Intake team is moving from stateless to stateful encoding and used Antithesis to hunt for protocol-breaking bugs, according to an Antithesis blog post dated 29 September. Datadog collects more than 100 trillion events per day, and the team said the change should cut transmitted data. Joy Zhang, a senior staff engineer, said the failures Antithesis surfaced "made us reevaluate how we were testing the system".

Smaller builds are following the same pattern. Developer Alex Grinman published a walkthrough on 29 September of Tinyboard, a Rust platform for ESP32 e-paper gadgets that wakes on a timer or button, draws once, and returns to deep sleep at about 10 µA. His platform clamps sleep to between 5 seconds and 24 hours so a bad deadline becomes a pause, not a reboot loop.

Put together, the week's announcements point one way: whether the boundary is a national firewall, an enterprise contract or a sandboxed agent, platforms are being asked to show where the limit is and who enforces it.