Yesterday, Anthropic teamed up with NVIDIA to release a security solution for enterprise-grade Agents, combining Anthropic's managed service Claude Managed Agents with NVIDIA's open-source runtime security system OpenShell.
The core idea is very straightforward: pull security controls out of the model's internals and into the external runtime environment. The specifics are as follows:
When the model is running tasks, passwords and access keys are all stored in a separate credential vault; the scheduling loop and the execution sandbox are completely split apart, so the Agent itself never comes into contact with real sensitive credentials from start to finish.
NVIDIA's open-source OpenShell serves as the external runtime control tool, intercepting all operations by default; only files, network connections, and tools with explicitly configured rules are allowed access. To prevent gaps in the rules, OpenShell also ships with a policy prover that uses formal mathematical proofs to verify the Agent's actual permission boundaries.
When running long tasks, if a network disconnection occurs, task progress is still persistently retained, and the Agent can distribute subtasks itself to push forward in parallel. The official cases listed include Notion, Rakuten, and Asana, mainly running in R&D delivery, cross-department collaboration, and daily office scenarios.
Honestly, taken as a whole, there's nothing new here and it's no new mysticism—it's just an extremely traditional systems engineering security solution.
After going around in a circle, from letting AI solve AI security itself, it's back to wrapping a "Faraday cage" around the outside of the AI. Writing this, I thought of "Person of Interest"—if you haven't seen it, I highly recommend it; it's a classic.