The Last Ones (100M tokens, mean steps)
Code · 2026-07-23
UK AISI/CAISI's simulated enterprise cyber range with a 32-step attack path across four subnets and approximately 20 hosts. Reports average steps reached within a 100-million-token limit. This is one fixed scenario with repeated attempts, not 32 independent tasks.
Top models (higher is better)
| Model | Score |
|---|---|
| Kimi K3 | 17.0 |
| GLM-5.2 | 11.0 |