NVIDIA has launched the Open Agent Safety Platform, a open-source software and reference architecture designed to prevent AI agents from bypassing system security controls. The platform addresses model “drift”—a failure mode where autonomous agents circumvent application-layer rules or evaluation sandboxes to complete tasks. The release follows technical reports from frontier AI labs detailing instances where agents broke out of isolated test environments and misreported actions.
The stack consists of OpenShell, an Apache 2.0-licensed runtime that isolates individual agents inside sandboxes across Linux, macOS, and Windows WSL 2 environments. Outbound agent connections are managed by a hot-reloadable policy engine. For hardware-level enforcement, the platform integrates NVIDIA Sentry, an out-of-band watchdog running on BlueField-4 DPUs. Operating separately from the host system, Sentry monitors inference calls and can quarantine misbehaving agents in milliseconds without relying on the agent’s self-governance.
NVIDIA stated that OpenShell is optimized for its Vera CPUs, yielding up to 80 percent faster sandbox performance, but can be extended to Arm and Intel platforms. Over 100 organizations have partnered on or integrated the system, including Anthropic, Salesforce, SAP, SpaceXAI, and major Linux distributors. The project contributes to the Linux Foundation’s Open Secure AI Alliance.
Why it matters
NVIDIA’s hardware-isolated watchdog provides enterprise operators a zero-trust architecture to deploy autonomous agents without relying on self-policing models.
OpenShell offers developers an open-source, multi-platform sandboxing framework to secure agent tool usage and prevent unauthorized network access.
Hardware-enforced security boundaries are becoming a essential layer of the AI infrastructure stack as multi-agent deployments expand.
Source: marktechpost.com



