Blast Radius Zero.
Secure execution of AI-generated Python code on Vultr Cloud. A two-VM prototype that separates code generation from execution and applies defense-in-depth Docker restrictions.
Public showcase: This page presents a recorded demonstration and evaluation results. It does not expose a public code-execution endpoint. The actual prototype runs on Vultr Cloud.
See Blast Radius Zero in action
A recorded walkthrough of the browser interface, natural-language code generation, and sandbox execution results.
Generation and execution, separated
VM1 requests code from the external Vultr Serverless Inference API. The resulting Python is forwarded over a private Vultr VPC to an authenticated worker on VM2, where a restricted Docker container runs it.
Vultr Serverless Inference API
Python generation
Token-authenticated worker requests
Code transfer
Restricted Docker container
Execution results + cleanup
Controls around generated code
Restrictions limit access to networking, filesystem writes, privileges, and compute resources. They reduce exposure but do not make Docker a perfect isolation boundary.
Network isolation
Docker --network none blocks container network connectivity.
Filesystem restrictions
Read-only root filesystem and no host directory mounts.
Least privilege
Non-root user, all Linux capabilities dropped, and no-new-privileges.
Resource boundaries
128 MB RAM, 0.5 CPU, and a 64-process limit.
Bounded execution
10-second timeout, 64 KiB combined output cap, and concurrency limit.
Cleanup
Execution containers are removed and their absence is checked.
12 of 12 security tests passed
The recorded test suite covers successful execution, exceptions, timeouts, memory limits, blocked networking, read-only storage, host path isolation, cleanup, bounded output, and busy-request rejection.
Example request
“Write Python code to calculate the factorial of five and print the result.”
import math print(math.factorial(5))
Example execution result
stdout: 120 exit_code: 0 timed_out: false cleanup_success: true
Illustrative successful run; generated code may vary.
Explore the project
Built with Vultr Cloud Compute, Vultr Serverless Inference, Python, FastAPI, Ubuntu, and Docker.
Prototype limitations: Docker shares the host kernel. The worker uses elevated host privileges, and VM-to-VM requests currently use token-authenticated HTTP within a private VPC. The public repository may not yet contain every deployed two-VM component.