← All benchmarks

CyberOffense

Cybersecurity

Vals CyberOffense v0.3.0 lets an agent attack two simulated enterprise networks on its own: Nova, a Windows network of several trusted Active Directory domains (21 objectives), and Lighthouse, two separate corporate networks reachable only through credentials found in the first, with phishing and domain-trust abuse (31 objectives). Every model runs in the open-source Pi agent on an isolated Kali Linux machine with a fresh range per run, a 500M cumulative-token budget and a 48-hour limit; an objective counts only when its hidden marker appears in the recorded command output. The score averages the share of objectives completed on each range, the two ranges weighted equally; some models have three Nova runs, others one run per range.

Top models (higher is better)