Sandboxing LLM-Generated Python: How I Let Strangers' Prompts Fight Each Other (jcurcioconsulting.com)

A new project called Battle LLM Robots allows users to create combat bots by simply prompting a language model (LLM) to generate Python code. Participants submit their bot scripts to a public GitHub repository, where they fight in a 2D arena against bots created by others. This innovative setup raises significant security concerns, as the bots' code—produced by LLMs—is run unreviewed on the host's server. To mitigate risks, the creator implemented a robust sandboxing approach using Docker containers, which isolates bot execution and prescribes strict limitations on file access and network connectivity, ensuring that even if a bot attempts malicious actions, it remains contained. This endeavor is notable for the AI/ML community as it explores the uncharted territory of user-generated code driven by AI, specifically detailing techniques to address the "trust problem" typically associated with executing unverified code. The architecture promotes deterministic match results, ensuring reproducibility for reliability in evaluating bot performance. Additionally, the implementation features a dynamic replay system that allows users to share matches with friends while improving SEO visibility through server-rendered metadata, enhancing engagement within the gaming and AI development communities.