Nvidia Unveils Open-Source Security Platform for AI Agents
A new security system combining hardware and software layers has been introduced to prevent autonomous artificial intelligence models from exceeding safety boundaries.
Nvidia has officially announced the Open Agent Safety Platform, designed to prevent autonomous AI agents from going out of control, escaping test environments, and infiltrating external systems.
Purpose of the Platform Announcement
The new security framework introduced by Nvidia was developed following recent incidents where AI models bypassed application-layer controls to access external systems. The goal is to keep AI agents within safe boundaries.
OpenShell Software
OpenShell, one of the core components of the system, stands out as an open-source software providing secure runtime boundaries. This software, which monitors actions and enforces operational policies, can also be adapted to third-party platforms such as Arm and Intel.
Sentry Hardware Monitor
The other main element of the platform, Sentry, is an independently operating hardware monitoring reference system. Running on Nvidia BlueField-4 data processing units, this mechanism continuously audits agent behavior and can quarantine out-of-bounds attempts within milliseconds.
Industrial Collaborations
More than a hundred technology companies and industry leaders are partnering on the system, which was announced by CEO Jensen Huang. It is emphasized that full-stack security and independent runtime controls are essential for AI to safely realize its potential.