Crucible

Crucible is an autonomous cyber range for testing AI defenders safely. A red AI (Muse Spark) probes a generated company network, a blue AI (Muse Spark Open, an open weight model on Baseten) contains the intrusion, and an independent auditor (Muse Wonder) checks every decision against the event log. You can take over the defender by voice, human runs are flagged and kept out of the benchmarks, and every debrief claim cites an event ID. On 20 held-out networks, blue contained all 18 intrusions with no protected file read when its sensors worked, and only 4 of 19 when sensors were off. A full exercise costs about 66 cents. Everything was built between Oct 26 and Nov 3, 2026, and the commit history shows it.