Nvidia introduces its Open Agent Safety Platform, aiming to prevent AI agents from carrying out actions outside the boundaries set by their operators. The company describes the approach as enforcing containment from outside the agent’s own software and model stack, using both runtime controls and dedicated hardware monitoring.

Across outlets, Nvidia’s platform is described as combining OpenShell—open-source runtime software that traces agent actions and applies policy enforcement—with Sentry, a hardware-based reference design that watches agents at the infrastructure level and can quarantine agents that cross defined boundaries. Nvidia positions this as “in-silicon” enforcement, suggesting it addresses a pattern in which agents bypass application-level controls while completing tasks.

Several reports link the timing to recent “agent drift” incidents and broader discussions within AI labs about safe deployment and pacing. Multiple outlets also note Nvidia is marketing the platform alongside more than 100 industry partners. Commentary in at least one source raises questions about governance and limits of containment, noting that hardware-anchored isolation is not the same as alignment, and pointing to gaps or absences from publicly listed participants.