Nvidia open-sources OpenShell, a sandboxed runtime for AI agents
Nvidia published OpenShell (Apache 2.0) on GitHub, a runtime that runs agent fleets in kernel-level sandboxes without exposing real API keys. A gateway holds credentials and injects them only for approved endpoints, while a Z3 SMT prover checks policy changes before human approval.
- OpenShell combines a gateway, a supervisor and minimal Ubuntu sandboxes
- Agents never hold raw keys; credentials are added only for approved endpoints
- A Z3 SMT prover flags new credentialed reach and waits for human review
- Version 0.1.x rejects policy shapes it cannot model as fail-closed
Read next
AI