We’re exploring specialist models for security work and building tools to investigate findings and test control over AI workloads.
Lumi · proposed frontier micro-model research
Lumi asks whether a small local specialist model can rank the source locations relevant to an existing security finding materially better than a cheap lexical baseline. The model programme is proposed; the public Trace workflow is working now.
Read the programme Start with Lumi TraceEggcracker
Eggcracker explores control and containment of unauthorised AI. Its documented public probe tests the mechanism with harmless sleeping processes—not a real AI workload. A separate historical recorded test is replayed on the demo page; nothing runs on your device.
See the Eggcracker probe Watch the demoInteractive exercises
Browser simulation · evidence-reading educational exercise
Choose the strongest claim that fictional evidence supports.
Can you prove the AI is stopped?Fixed synthetic browser educational exercise
Review provenance, scope, correction, rejection and forgetting in one synthetic example.
Memory Control Lab