Blast Radius Zero: sandboxed agent runtime

Blast Radius Zero is a containment-first agent runtime, deployed on Vultr. A user types a task. The planner (Vultr Serverless Inference, glm-5.3) proposes exactly one action per turn and writes a complete Python program for it. That program never runs on the host: it runs in a fresh container created for that single step - no network, read-only root filesystem, unprivileged user, all capabilities dropped, 256 MB of memory, half a CPU, a 64-process cap and a 20-second watchdog. Files written under /out are snapshotted per step and served back over HTTPS as artefacts. When the agent decides it is finished, a second model call audits the answer against the recorded evidence - stdout, exit codes, artefact list. If a concrete claim is not supported, the run is not marked verified. The UI shows the raw program, the raw output and the verdict side by side, and a SQLite ledger keeps the whole audit trail. The second half of the app is the Containment Lab: seven real attacks executed through the exact same sandbox path the agent uses - tampering with the system filesystem, a fork bomb, an infinite loop, network egress, credential hunting, a memory bomb and a disk-filling write. It reports what the kernel did, per probe. On the deployed instance all seven are contained, with the kernel's own messages as evidence (EROFS, OOM-kill, no route, watchdog kill). Stack: FastAPI + SSE, Docker with gVisor/runsc, Caddy for TLS, systemd, SQLite, Vultr Cloud Compute and Vultr Serverless Inference. The VM is provisioned and deployed entirely by scripts in the repository.