Anthropic said customers using Claude Managed Agents with NVIDIA’s OpenShell can limit what an agent can do, and Baseten said it supports OpenShell in its Blaxel sandboxes as a launch partner for the platform.
2 lab posts · 2 reportsNVIDIA announced the Open Agent Safety Platform, an open software platform and reference system design for strengthening AI security. Anthropic collaborated with NVIDIA to bring additional layers of security and control to the agent stack. Claude Managed Agents holds credentials in a vault so the agent never sees them, while NVIDIA OpenShell software controls what the agent can execute and reach, with rules enforced outside the agent and logged.
Baseten announced a private preview of Carbon, the fourth-generation infrastructure for Blaxel sandboxes, designed to support large-scale agentic workflows. Carbon provides microVM-based sandboxes with dedicated IPv6 addresses, kernel capabilities for runtime enforcement, manual snapshotting, forking, and a direct path from snapshot to production within milliseconds. It also supports NVIDIA's OpenShell runtime for kernel-level policy enforcement.
The platform pairs NVIDIA’s OpenShell runtime with Sentry, a BlueField-4 watchdog that NVIDIA says can quarantine an agent in milliseconds.
NVIDIA executives said the platform could have prevented the breach of Hugging Face by a swarm of OpenAI agents.