Sandboxed AI evaluation and execution service for untrusted code
Build a service that uses AI agents to evaluate submitted projects or execute agent-generated code inside isolated, disposable sandboxes rather than on a judge's or developer's local machine.
From Google Cloud Tech — How to build and scale multi-agent AI systems on GKE at 32:27
Problem: Downloading and running unknown code can cause code injection, resource abuse, container escape, data exposure, or damage to production systems. Human evaluation also produces inconsistent judgments when reviewers have different amounts of time and apply personal biases.
For: Hackathon organizers, engineering teams evaluating submitted code, and organizations running agents that generate or execute code against infrastructure.
Examples
- The GKE 10 year hackathon produced 400-something project submissions, motivating an automated baseline for consistent evaluation.