AegisBox is a containment-first autonomous execution agent designed for Vultr Challenge 1: Blast Radius Zero.
Key Architecture:
- Plans multi-step execution tasks using Vultr Serverless Inference (meta/llama-3.3-70b-instruct).
- Dispatches untrusted code into ephemeral, unprivileged sandboxes hosted on a Vultr Cloud Compute VM (IP: 144.202.31.205).
- Enforces strict secret hygiene: host environment variables and Vultr API keys are completely isolated from the execution runtime.
- Mints deterministic SHA-256 cryptographic receipts for every run.
- Demonstrates active containment moments by intercepting unauthorized network egress and system reconnaissance.
GitHub: https://github.com/Modeste01/aegisbox
The architecture screenshot is at: https://github.com/Modeste01/aegisbox/blob/master/architecture_diagram.png
Tech Stack:
• Cloud Host: Vultr Cloud Compute (Ubuntu 24.04 VM at 144.202.31.205)
• AI & LLM: Vultr Serverless Inference (api.vultrinference.com/v1, Meta Llama 3.3 70B)
• Backend: Python 3, FastAPI, Uvicorn, Pydantic
• Sandbox & Containment: Ephemeral workspace isolation, runtime socket air-gap, 5s watchdog ceiling
• Verification: Cryptographic SHA-256 receipt minting
• Frontend: Responsive Web Console (HTML5, Tailwind CSS, JetBrains Mono)
• Deployment: Docker & Docker Compose