Products

Proposed frontier micro-model research

Lumi

Can a small local specialist model trace a security finding to the code involved?

Lumi is noqt’s proposed frontier micro-model research programme for security localisation. No Lumi micro-model has been trained, qualified or shown to beat the current baseline. Lumi Trace is the working public tool available today.

Supply a finding and a local Python repository or supported archive. Trace ranks likely source paths and produces JSON or SARIF, a standard format for static-analysis results. The repository stays local.

In the walkthrough, Bandit performs the synthetic scan and supplies the finding; Lumi checks that supplied result. Lumi is not a vulnerability scanner: it doesn’t discover another issue or prove a vulnerability exists, and may return no confident match.

Requires CPython 3.11 or 3.12. Docker is optional and used only for reproduction. Use source code you are authorised to inspect. Follow the README for supported inputs and setup.

Documented probe

Lumi Eggcracker

Inspect a containment mechanism on a disposable Linux host.

Eggcracker’s aim is to help people retain control over unauthorised AI workloads. The public starting point is narrower: a documented probe that starts two harmless sleeping processes and tests containment.

The Hugging Face Space contains an older source snapshot. It does not run Eggcracker in the browser. Use the pinned GitHub README above for this probe.

A separate historical recorded test covers one captured supported workload and a separate test task. It is replayed only; nothing runs on your device.