Nvidia builds hardware boundaries for AI agents
Nvidia's new Open Agent Safety Platform pushes agent containment down into hardware — OpenShell enforces CPU-level boundaries and Sentry quarantines a compromised host in milliseconds, with Anthropic and 100+ other partners already signed on. It's a sign the industry no longer trusts software-only sandboxing to catch what a fast-moving agent might do, which matters if you think about agent safety as an engineering problem rather than a policy one.
What’s hardware-level agent containment?
Most agent safety today is enforced in software: permission prompts, sandboxed processes, network allowlists, and prompt-level guardrails. All of these run on the same CPU as the agent itself, which means a sufficiently capable or exploited agent can, in principle, interfere with the very mechanism meant to contain it.
Hardware-level containment moves some of that enforcement below the operating system: a chip or co-processor sets boundaries — say, which memory regions or compute quotas an agent process can touch — that the agent's own code cannot rewrite, because it never has execution access to that layer at all.
Nvidia's pitch is that this catches what software sandboxing misses in a crisis: a runaway or compromised agent can be quarantined in milliseconds by a watchdog sitting outside its own blast radius, rather than waiting for a monitoring script running on the same box to notice and react.
The open question is whether this becomes a real industry-wide safety layer or mostly a way for Nvidia to sell more silicon under a safety banner — hardware containment adds a new dependency and a new vendor to trust, and it's a second line of defense on top of good software design, not a replacement for it.
via cnbc