Nvidia Proposes Dedicated Safety Chip Beside Every AI Agent
Nvidia's OpenShell platform would give every autonomous AI agent a separate monitoring chip that watches its actions from outside the model, an architecture the company says agents cannot rewrite.
Nvidia is proposing that every autonomous AI agent run alongside a dedicated monitoring chip designed to cut off rogue behavior before it spreads. The effort, described in a blog post under the OpenShell banner and dubbed the Open Agent Safety Platform, pairs a running agent with a companion watchdog called Sentry that observes the agent's actions from outside the model itself. Nvidia frames the chip as a hardware root of trust for a problem the software industry has not solved: what to do when an agent goes off script.
The architecture flips the usual safety setup. Rather than asking a model to police itself through prompts or a second model instance, Sentry sits on dedicated silicon and monitors outputs, tool calls and resource requests in real time. Nvidia's argument is that an agent which has been compromised, misaligned or pushed into an unintended loop should be stopped by hardware it cannot rewrite. That distinction matters because most guardrails today run in the same process, or at least the same trust boundary, as the agent they are meant to contain.
Nvidia is positioning OpenShell as an open platform rather than a proprietary Nvidia-only stack, which would let cloud providers and enterprise buyers bolt on a safety layer without rebuilding their agent runtimes. The timing tracks the shift of agent frameworks from OpenAI, Anthropic, Google and a long tail of startups out of demos and into production, where a single runaway agent can reach databases, payment rails and internal APIs. Nvidia has spent years adding confidential computing features to its data-center GPUs, and Sentry extends that pattern from keys and attestation to agent behavior.
Skeptics have already raised the central objection: a watchdog chip only helps if the agent's traffic actually passes through it, and many deployments call external tools directly over the network. Nvidia has not published latency figures, power budgets or pricing for Sentry, nor explained how it handles agents that spawn sub-agents or delegate work to third-party services. Those gaps leave the proposal split between those who see a sensible hardware root of trust and those who see another layer that can be routed around.
There is precedent for the idea. Hardware security modules and trusted platform modules already enforce rules for cryptographic keys and device attestation. Extending that model to agent behavior is a larger claim, because 'rogue' is far harder to define than 'unauthorized key access.' A chip can enforce a rule; it cannot decide what the rule should be, and that policy question is where most agent safety work remains stuck.
What to watch next is whether Nvidia ships OpenShell as an open specification other chipmakers can implement, or keeps Sentry tied to its own hardware. The company has also not said when developers can test it or which agent frameworks will be supported at launch. Until those details land, the proposal reads as a credible direction with an unproven enforcement story.