SandboxOps: Blast Radius Zero for AI Agents

SandboxOps lets teams give an AI agent real work without giving it a real shell. Every task follows the same loop: plan, sandbox, verify, evidence. A user describes a task in plain language, Vultr Serverless Inference (model glm-5.3) turns it into a short Python script, and that script runs inside a throwaway Docker container on a Vultr Cloud Compute VM (Chicago, Ubuntu 24.04) with no network, a read-only filesystem, a non-root user, all Linux capabilities dropped, and hard CPU, memory and process limits. The container is destroyed after each run. The output is checked against a verification rule when one exists, and every run produces an evidence record with the exact code, stdout, stderr, exit code and a SHA-256 hash, so a reviewer can audit what the agent actually did. Live on Vultr today: the bill task verifies at exactly $551.00; a free-form prompt "sum of first 50 primes" returns 5117 via glm-5.3; a network exfiltration attempt is CONTAINED; an infinite loop is CONTAINED by the sandbox timeout. Health endpoint reports backend=docker and planner=vultr-inference. The target buyer is finance and healthcare operations teams that want LLM-generated reconciliations and bill checks without risking their servers or data. The app is a small web service with a /health endpoint, deployed with one script and run by systemd. It is MIT licensed, with automated tests for the sandbox and the web API. Honest limits: free-form prompts are not auto-verified; containers still share the VM kernel (gVisor is next).