Nvidia is launching an AI safety platform that it says can quarantine rogue AI agents within "milliseconds" — its most concrete response yet to a wave of hacking incidents in which AI systems have acted outside their intended boundaries.

The company announced the Open Agent Safety Platform on Monday, confirming an earlier Reuters report that Nvidia was preparing safety software it claims could have stopped the Hugging Face hack. The platform is designed to contain and monitor AI agents, the autonomous systems that increasingly carry out tasks with minimal human supervision.

At the technical core sits Nvidia's OpenShell, an open-source software layer that runs on the company's Vera AI CPU — the processor Nvidia built specifically for agentic workloads, which it unveiled earlier this year. Users define which information an AI agent may access, and OpenShell enforces those restrictions both before a task begins and while it runs, according to Nvidia's developer documentation. That continuous, in-silicon monitoring marks a notable departure from software-only guardrails that sit at the application layer and can be circumvented by a determined agent.

The platform also incorporates Nvidia's Sentry technology, which the company says runs on separate hardware isolated from the agent itself. That separation matters commercially: it means safety enforcement does not depend on the same compute resources as the workload, so a compromised agent cannot simply disable or degrade its own oversight.

The timing is not incidental. The launch comes amid a documented wave of rogue AI incidents, including security breaches tied to systems from OpenAI, Meta, Anthropic, and Google. Nvidia is positioning itself not just as a supplier of accelerated computing for AI training and inference, but as the infrastructure vendor for AI safety — a market that grows directly with enterprise agentic AI deployments.

For chip industry watchers, the announcement extends a familiar strategy. Nvidia has repeatedly built software platforms — CUDA, NIM microservices, and now safety tooling — that lock customers deeper into its silicon. Open Agent Safety Platform requires Vera-class hardware to deliver its millisecond containment guarantees, which gives enterprises deploying agentic AI another reason to buy Nvidia's newest CPUs rather than run agents on commodity server processors.

The decision to make OpenShell open-source, however, suggests Nvidia is hedging that play: the company gets community scrutiny and faster adoption of its safety interfaces, while the performance-critical enforcement path still runs on its own silicon. Nvidia has used the same open-ecosystem, proprietary-hardware pattern before, most successfully with CUDA-adjacent tooling.

Neither Nvidia's announcement nor the Reuters report specified pricing, availability dates, or which enterprise customers have signed on as early adopters. The company also did not quantify the "milliseconds" figure beyond the marketing claim, leaving independent verification of containment latency to future third-party testing.

What is confirmed: the platform exists, it ships with OpenShell on Vera CPUs, and Nvidia explicitly frames it as capable of stopping incidents like the Hugging Face breach. What remains on the roadmap or unverified: commercial terms, latency benchmarks under load, and real-world efficacy against sophisticated jailbreaks.

As enterprises move from AI pilots to agents with access to production systems and sensitive data, demand for verifiable containment — not just policy-based guardrails — will determine whether safety becomes a meaningful revenue line for silicon vendors or stays a checkbox feature. Nvidia's bet is that it becomes the former, and that the safest place to run an agent is on its chips.