ProofBox AI

ProofBox AI is a security-first coding agent for people who want AI to write and run code without putting their own machine or data at risk. A user describes a task in plain English. Ollama Cloud writes Python code for it. Before anything runs, ProofBox AI checks the code statically, reading it for syntax errors, dangerous functions, and risky imports, without executing it. The user then reviews the code and decides whether to run it. Nothing executes automatically. When the user clicks Run, the code is sent over SSH to a Vultr Cloud Compute instance dedicated to execution. It runs as an unprivileged user with no sudo access, no access to the host’s secrets, and no internet connection. Firewall rules block all outbound traffic from that user, and the setup script tests the block before the server is used. Each run is limited in time, memory, output size, and the number of processes it can start, and its temporary files are deleted afterwards. The result reports the output, the exit code, the run time, and whether cleanup was confirmed. ProofBox AI also scans public GitHub repositories for hardcoded secrets and dangerous patterns. It pins each scan to a specific commit, redacts any credentials it finds, and streams findings live. Every generation, check, scan, and sandbox run is saved to an audit history that can be exported as JSON or SARIF, the format used by GitHub’s code scanning. A readiness check runs live against the sandbox. The dashboard only shows it as ready when the connection works, cleanup is confirmed, and the sandbox cannot reach the internet. The project is built with Python, FastAPI, SQLite, and a plain HTML and JavaScript dashboard. Static analysis is pattern-based, so it can miss issues and flag things that are safe. A clean result is not proof that code is secure.