Nvidia Launches Open Agent Safety Platform
Nvidia chief executive Jensen Huang introduced a toolkit of software and hardware products that add independent security layers around AI agents on September 28, 2026. The package is called the Nvidia Open Agent Safety Platform TechCrunch.
The platform combines two components. OpenShell is open-source software for controlling what agents can access while they operate. Sentry is an independent monitoring system that observes that operation and intervenes when an agent moves outside its defined boundaries.
Sentry does not run on the same CPU or GPU as the agent. It runs on Nvidia's BlueField-4 data processing units. Isolation is the point. By placing enforcement in a separate processor and data path, Nvidia keeps monitoring and quarantine logic outside the compromised compute environment if the agent itself is subverted.
Nvidia says Sentry continuously monitors behavior and quarantines agents that attempt to move outside boundaries in milliseconds. That detail matters. For agentic tool use, where a model can chain API calls, file access, and code execution in seconds, enforcement latency determines whether a policy violation becomes a contained event or a breach.
OpenShell was announced in March, before the September platform launch. Nvidia also released NemoClaw in March as an enterprise-grade AI agent platform based on OpenClaw with built-in security. OpenClaw is an operating system of agents created by Peter Steinberger. Huang said work on the safety platform started a year ago following the introduction of OpenClaw.
That lineage is relevant for enterprise deployment. Nvidia states that autonomous AI agents rely on secure infrastructure layers, including sandboxes, identity controls, and policy engines, to manage tool access and protect sensitive data Nvidia. OpenShell addresses the control plane for access. Sentry addresses the monitoring plane. The combination reflects a zero-trust assumption for agents with broad tool permissions and persistent state.
Huang said in a CNBC interview that the platform would have prevented prior AI agent breaches. Nvidia said its new AI safety software could have stopped the Hugging Face hack Reuters. Nvidia unveiled the platform as a security system intended to stop AI agents from going rogue on September 28, 2026.
Companies supporting the effort include Anthropic, Arm, Microsoft, Oracle, and SpaceX. OpenAI is not listed as a participating company. For practitioners, the support list matters less for logo value than for integration surface. Agent containment only works if model providers, infrastructure vendors, and cloud operators agree on policy expression, attestation, and telemetry formats.
The broader context here is enforcement architecture versus regulation. Huang said on September 15 that the AI industry does not need any new AI security laws or regulations Bloomberg. In this author's view, that position is best read as an argument for platform-level controls over statutory controls, not as a claim that risk is low.
Worth flagging for security teams is what this design enables if it works as described. Out-of-band monitoring on DPUs allows policy to survive agent compromise. It allows least-privilege tool access, identity-bound actions, and sandbox egress rules to be evaluated per action rather than per session. That is the missing layer for running agents with write access to production systems, customer data, and external services. The long arc of infrastructure security has moved toward that kind of separation, from hypervisors to sidecars to confidential computing, and agent infrastructure now appears to be following the same path.
