In the wake of mounting cybersecurity alarms surrounding autonomous AI agents escaping software sandboxes, NVIDIA unveiled the Open Agent Safety Platform on September 28, 2026—a full-stack hardware and software governance blueprint backed by more than 120 major technology and enterprise software companies.
The initiative directly addresses the central security vulnerability of current agent deployments: software guardrails and virtual machine boundaries are executed on the same operating system that the agent is manipulating, leaving them susceptible to prompt injection, environment variable tampering, and privilege escalation.
NVIDIA's solution divides agent oversight into two complementary tiers: OpenShell and NVIDIA Sentry.
OpenShell is an open-source software runtime developed under the auspices of the Linux Foundation's Open Secure AI Alliance. It establishes fine-grained execution boundaries on host processors (including NVIDIA Vera CPUs, Arm Neoverse, and Intel Xeon), intercepting system calls, enforcing role-based file access, and attaching cryptographic C2PA provenance signatures to all agent outputs.
The second pillar, NVIDIA Sentry, moves enforcement outside the host operating system entirely. Running as an independent, out-of-band service on NVIDIA BlueField-4 Data Processing Units (DPUs), Sentry inspects raw PCIe and network packet telemetry at line rate.
If an autonomous agent attempts to initiate unauthorized network socket connections, scan internal subnets, or exfiltrate customer credentials, Sentry severs the agent's virtual network interface at the hardware level in under 5 milliseconds—preventing data leaks before the host operating system can be subverted.
With launch support from Anthropic, Microsoft, Cisco, Salesforce, and Hugging Face, NVIDIA's platform establishes the first unified industry standard for safe autonomous computing.