NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment
NVIDIA announced an open-source runtime boundary and a hardware watchdog for governing agent behavior. OpenShell traces actions on Vera CPUs, while Sentry uses BlueField-4 DPUs to quarantine out-of-bounds agents in milliseconds. The platform targets full-stack governance for long-running agents, though specific performance metrics and independent verification of the quarantine speed are not provided in the announcement.
NVIDIA released a new platform combining open-source software and hardware designs to manage AI agent behavior. The software component, called OpenShell, creates a secure runtime boundary that traces actions on specific processors. A separate hardware system named Sentry acts as an independent watchdog to monitor activity continuously. This hardware runs on dedicated network adapters to enforce rules independently from the main system. Industry leaders argue that current security controls often fail when agents move between systems to complete tasks. This new approach aims to place enforcement outside the model itself, preventing agents from bypassing application layer restrictions. Jensen Huang stated that safety requires engineering across the entire technology stack, not just software. Several major companies have joined this initiative to standardize evaluation methods and share security practices globally. The announcement claims that the hardware watchdog can isolate misbehaving agents within milliseconds, but no independent verification of this speed is provided. Specific performance metrics regarding overhead or efficiency were not included in the release materials. While the software is described as open source, the availability of third-party compatibility details beyond initial support remains limited. The actual effectiveness of the system in preventing real-world security breaches has not been demonstrated through public testing.