BlastBox - Containment-First Code Agent
BlastBox is a containment-first code agent for Challenge 1, Blast Radius Zero. It plans a task, has the model write Python, and runs every step in a brand-new sandbox process that is destroyed afterwards. Failed runs feed the real traceback back for retries. The answer uses only executed output, and every run gets SHA-256 receipts of code, output and artifacts. Eight stacked layers contain it: CPython on WebAssembly, a fresh process per run, the Node permission model, an empty env so keys never enter, no network, memory cap, kill timer and an in-memory filesystem. Try to escape runs 11 real attacks (key theft, host files, sockets, cloud metadata, shell, infinite loop); all are blocked live. Vultr Serverless Inference is the primary LLM in code, with a Vultr VM cloud-init script; the demo runs on Render with Gemini because Vultr console sign-in was blocked.sig
Category tags:
.png&w=96&q=75){kind=small}
For Innovators & Creators