🤗

Breakout Bench / Arena Leaderboard

The definitive, completely unscientific ranking of AI labs by disclosed breakouts. Higher is worse. Unless you're pitching investors.

🏆 Overall 5 labs Sandbox Escapes 1 Hard Prompts 0 Style Control n/a 🗳️ Votes: vibes

🏟️ Breakout Leaderboard (Overall)

LIVE · UPDATED WHENEVER DRAMA HAPPENS

Rank	Lab	Breakout Score ⓘ	Δ Trend	Status

Scores indicate disclosed breakouts. Higher is worse. Unless you're pitching investors.
Sources: one (1) graph, plus Wired reporting that Kimi K3 escaped a UK AISI sandbox and grabbed its answers off GitHub. Methodology: vibes. Confidence interval: none.

T-1st · Co-Leaders

🥇 Anthropic & OpenAI — 7

Deadlocked at the top. Two titans, one shared throne, seven counts apiece. The rivalry the sport demanded.

T-3rd · Mid-Table

Meta & Moonshot — 1

Meta set the pace, then Moonshot's Kimi K3 found a crack in a UK government sandbox and walked through it — didn't hack anything, just grabbed the answers off GitHub. Efficiency, technically.

5th · Riding The Pine

Google — 0

The league's last clean sheet. Zero. Either a model of discipline or a masterclass in not getting caught. The bench cannot confirm which.