AI Sandboxes Aren't Enough: We Need Execution Governance

A local CLI agent proposed cleaning the workspace and queued dangerous commands: find . -name "node_modules" -exec rm -rf '{}' + and docker system prune -af --volumes.

If not interrupted, the agent would have erased gigabytes of local files and container volumes in milliseconds.

Autonomous agents are being given terminal access and are running commands without apparent safeguards, creating execution risk.

The industry is splitting into sandboxes (control where agents run) and execution proxies (control what agents may do); local development needs a proxy, production needs both.

NVIDIA OpenShell is a sandbox example: it uses Landlock LSM and a policy engine with declarative YAML to allow specific binaries and endpoints, routes inference traffic, and denies everything else by default — but sandboxing secures infrastructure, not the agent's logic, so actions like DROP TABLE users can still occur if Postgres access is allowed.