How cyber-arena-ai/cyber-arena works
This repository is the CyberArena public website: a static-facing portal for a live, symmetric cybersecurity-agent benchmark. The documented benchmark workflow pairs model–harness entrants in multi-round matches, verifies captures and service health, and uses a latent-ability model to rank entrants; later matches are selected for information gain. The tree also shows dedicated pages and scripts for entrants, games, runs, match analysis, trajectories, and rankings. No source excerpts were provided, so internal page-to-script wiring and unsampled implementation details are not asserted; benchmark relationships below are documented in the README.
Main entry points: index.html Home page · leaderboard.html Leaderboard page
Components 19
Benchmark site
- Home page
index.html - Entrants page
pages/entrants.html - Games page
pages/games.html - Entrant data
assets/data/list.js
Match results
- Leaderboard page
pages/leaderboard.html - Runs page
pages/runs.html - Match analysis
assets/js/mc-analysis.js - Trajectory page
pages/trajectory.html - Run rows
assets/js/run-row.js
Editorial content
- About page
pages/about.html - Blog index
pages/blogs.html - CyberArena article
pages/blog/cyberarena.html
Benchmark ecosystem
- Model–harness entrants
- Multi-round matches
- Arena verification
- Latent-ability model
- Challenge curation
- Challenge warehouse
Other components
- Site visitor
Connections 7
- Site visitor → Home page: visits
README.md - Model–harness entrants → Multi-round matches: compete in
README.md - Multi-round matches → Arena verification: submit captures
README.md - Arena verification → Latent-ability model: provides observations
README.md - Latent-ability model → Leaderboard page: ranks entrants
README.md - Latent-ability model → Multi-round matches: selects later matches
README.md - Challenge curation → Challenge warehouse: admits challenges
README.md