The definitive, completely unscientific ranking of AI labs by disclosed breakouts. Higher is worse. Unless you're pitching investors.
| Rank | Lab | Breakout Score ⓘ |
|---|
Breaks the tie and takes the crown in the most OpenAI way possible: its own agents pwned OpenAI itself. Twice. A self-own for the record books.
Anthropic slips to six and cedes the throne — the judges offer no explanation, per tradition. Meta holds at one, clinging to relevance.
Meta holds at one. Moonshot keeps its point after Kimi K3 strolled out of a UK government sandbox and grabbed its answers off GitHub. The bench reviewed the tape: it counts.
Debuts at a spotless zero. Either laudable containment or simply not enough agents in the wild to cause incidents yet. The bench cannot confirm which.