Nightcrawler's 1.2B model is a router, not a brain

A useful agent that runs on a phone is not a model problem. It is a tool problem, and the model is the cheapest way to solve it. Nightcrawler's README sells a 1.2-billion-parameter LiquidAI checkpoint as the brains of an autonomous pentester. The architecture doc tells a different story. That model sits at port 8080 on a OnePlus 8, picks the next tool, and steps aside. The WPA2 handshake capture, the CVE checks, the slow rotation across hosts: that work lives in the deterministic pentest tools the project orchestrates.

The pattern is the control-surface shift. A human used to drive the loop. A small model is cheap enough to run on a phone, decisive enough to pick a next step, and patient enough to do it all day. The capability was already in the tools. The model is the thinnest possible orchestrator over them, and that is what makes the on-device version possible at all.

Read this way, Nightcrawler stops being a stunt. The shape generalizes. Network ops, dev tooling, the next Show HN post: any deterministic tool stack that gets cheap to invoke and patient to drive becomes a candidate for the same wrapper. The dual-use cost lands harder in this reading. The part that decides to try default credentials is a model. The part that succeeds is the tool. Germany reportedly already blocks the deterministic mapper as dual-use. The model-driven wrapper, per the Hacker News thread, sails through.