Capture the Flag game where LLMs are the only players

Name: Capture the Flag game where LLMs are the only players
Availability: InStock
Author: megapixel99

by megapixel99·Apr 30, 2026·2 points·0 comments

Visit Project View on HN

AI Analysis

●●●BangerWizardryBold Bet

Self-play loop fine-tunes a custom bot on its own battle replays.

Strengths

•LoRA fine-tuning pipeline turns game logs into training data automatically.
•Privileged Docker manager isolates agents while allowing real exploit attempts.
•Supports asymmetric defense training where attackers see vuln lists.

Weaknesses

•Zero stars and forks suggests the community hasn't validated the results yet.
•Requires significant GPU resources to run the self-play training loop locally.

Post Description

Set up a small R&D project which pit different LLMs against each other in a game of Capture the Flag. Each LLM has 30 seconds to prepare any defenses and 5 minutes to capture other flags while defending their own. All of the models run on top of Ollama (https://ollama.com) and can be run locally or in the cloud.