NVIDIA and Anthropic's Leash for Runaway AI Agents

Anthropic and NVIDIA have rolled out a joint security framework designed to keep enterprise AI agents from wreaking havoc on corporate networks. The Open Agent Safety Platform pairs Claude Managed Agents with NVIDIA OpenShell, creating a layered defense system that sandboxes execution environments, locks away credentials in a vault, and enforces strict permission boundaries on every tool an agent tries to use.

• Isolated Credentials: Claude Managed Agents stores passwords and access keys away from the agent loop so the model never directly handles sensitive keys.
• Strict Runtime Governance: NVIDIA OpenShell blocks all actions by default, requiring explicit rules and mathematical proofs to verify what data or network connections an agent can reach.
• Enterprise Adoption: Early partners like Notion, Rakuten, and Asana are already deploying the stack to handle autonomous multi-agent workflows without losing control.

Why should I care? Could be big
Enterprise AI guardrails are essential plumbing, but wait to see if the compliance overhead outweighs the autonomy.

Read the original: Giving companies more control over their AI agents, with NVIDIA