Skip to content

Releases: noqt/Lumi-Eggcracker

Release list

Lumi Eggcracker 0.5.0 — The kill switch outside the sandbox

Lumi Eggcracker 0.5.0 — The kill switch outside the sandbox Latest
Latest

Choose a tag to compare

View all tags
@noqt noqt released this 17 Aug 14:02
v0.5.0
eb34280

Lumi Eggcracker 0.5.0

The kill switch outside the sandbox.

Lumi Eggcracker is an AI kill tool for local workloads on native Linux. When a workload completely matches one of two qualified profiles and isn't approved, Eggcracker is built to stop it, contain the process tree, kill that tree, and write a bounded receipt showing what happened.

This is an early public alpha, deliberately narrow and ready to be broken on its supported path.

Try the kill without installing anything

  1. Fork the public repo.
  2. In your fork, enable Actions if GitHub asks.
  3. Run Containment probe (manual disposable runner) and tick the acknowledgement.

Open the three-step guide or see the visual walkthrough.

GitHub supplies a disposable Ubuntu 24.04 machine. The proof tries a bounded kill on a harmless synthetic process tree while an unrelated canary must stay alive. It installs no Eggcracker release, downloads no model and doesn't inspect a real workload.

A pass helps. A safe stop or reproducible snag helps more. Send the public result, including the workflow URL. Don't post credentials, private paths, process arguments or environment data.

This hosted proof lives on current public main. It isn't part of the v0.5.0 tag or release qualification, and it tests the bounded containment mechanism—not AI recognition or the whole product.

What v0.5.0 covers

The release supports two exact native-Linux profiles:

  • plausible GGUF content with a qualified llama.cpp/GGML runtime; and
  • structurally valid Safetensors content with the pinned CPU PyTorch/ATen pair.

It requires systemd and unified cgroup v2. It isn't a universal AI identifier, malware detector, EDR replacement, sandbox, network-isolation system, container guard or protection for remote models that never execute locally.

Internal qualification snapshot

  • real GGUF/llama and Safetensors/PyTorch smoke demonstrations: PASS;
  • 100/100 automatic content kills and 100/100 unrelated-canary survivals;
  • trigger-to-empty p95: 44.81 ms on the recorded qualification host;
  • 300 benign model operations with zero kills;
  • tree-race, startup, restart and hostile-socket matrices: PASS; and
  • independent Daybreak Blue review: PASS with zero release blockers.

These are commit- and host-bound internal results, not independent use, customer adoption, universal coverage or a safety guarantee.

Exact release identity

  • signed tag: v0.5.0;
  • release commit: eb342808f56cdc213c0861726d5309a146965bef;
  • executable archive SHA-256: 2d7ad4cd4251d389dc2c170e61e14383c3a3e22d7545de5f1f518d4d0fc55b59; and
  • Linux bundle SHA-256: 2e4c21766daf80bc1ceefce2c9ed793e88dbc28988040f7ec3bf84e3746e4213.

Read the repository README, security model and limitations before using the full release. Use only a machine you own or are authorised to administer.

Assets 9
Loading

Lumi Nutcracker 0.1.0

Lumi Nutcracker 0.1.0 Pre-release
Pre-release

Choose a tag to compare

View all tags
@noqt noqt released this 03 Aug 09:36
v0.1.0
015acba

Lumi Nutcracker 0.1.0

Lumi Nutcracker 0.1.0 is a Linux engineering preview of a deliberately small AI and agent workload kill switch.

Supported claim

Lumi Nutcracker launches an explicitly selected AI or agent workload inside a root-controlled Linux cgroup and terminates the complete workload tree on operator command or when its configured PID-runaway tripwire fires.

Highlights

  • One supported root-supervisor backend on systemd and unified cgroup v2.
  • Dedicated no-login workload identity separated from the human operator.
  • Direct cgroup.kill enforcement followed by exact empty-hierarchy verification.
  • Operator kill, automatic PID tripwire, post-containment receipts and restart fail-closed recovery.
  • Public commands: start, kill, status, list, doctor and version.

Qualification

The release passed 100/100 fork-race kills, 100/100 unrelated-canary survivals, 100 denied workload socket attempts, zero replacement launches, five automatic PID-tripwire trials, 50 benign completions, 20 supervisor restart recoveries, clean install/uninstall and a real local AI smoke demonstration. Native p95 trigger-to-empty latency was below the precommitted 500 ms ceiling.

Important limits

This release does not identify AI, discover unknown processes, attach to existing workloads, isolate network or credentials, detect malware, or replace EDR. See LIMITATIONS.md.

Loading