Nightcrawler is an open source penetration testing agent that runs entirely on a smartphone. Place a OnePlus 8 with Kali NetHunter on a network, walk away, and the agent discovers devices, maps services, searches for vulnerabilities and generates a structured report. Everything local: the 1.2 billion parameter model (LFM2.5) runs on the phone’s GPU via OpenCL. No internet connection, no cloud APIs.

It’s the first concrete signal of a completely offline security agent on mobile devices. Those building privacy-first systems look at small models on edge devices as an alternative to sending data to servers. Nightcrawler shows that a 1.2B model can orchestrate a complex multi-step task: scanning, enumeration, exploitation, reporting. The agent works like a patient human pentester, one action per turn, with knowledge that accumulates gradually over hours of silent work.

As we reported on July 24, frontier model guardrails block offensive security researchers and push professionals toward open local models. Nightcrawler is the materialization of that trend: when the cloud closes the door, the phone is enough.

The repo is public on GitHub, version v0.1.0, and is explicitly declared as an experiment, not a finished product.

In detail

The context: scanners, humans, agents.

Traditional pentesting is done with automated scanners (nmap, Nessus) that bombard every host in seconds, or with professionals working by hand for hours. Nightcrawler sits in the middle: an agent that works like a patient pentester. One action per turn, knowledge that accumulates, hours of work. The README makes it clear: this makes it harder to detect than scanners that hit everything at once.

The stealth approach is not a technical detail. A traditional scanner generates traffic that a monitoring system (IDS/IPS) detects in seconds. An agent that takes one action per turn, waits, observes, then takes the next one, blends into the network noise. For those building defenses, this means realistic testing is no longer just the fast scanner: it must include the patient adversary case.

How it works.

The agent runs on a rooted OnePlus 8 with Kali NetHunter. The brain is LFM2.5-1.2B-Instruct-Heretic, a 1.2 billion parameter model from LiquidAI’s LFM2.5 family, the same one we saw in July focusing on efficiency. The model is loaded onto the phone’s GPU via OpenCL and exposes a local interface on port 8080.

The agent uses the Model Context Protocol (MCP) to call tools: network scanners, enumeration tools, exploits. A local web dashboard serves as Command and Control to monitor and guide the agent as it works. The workflow has five phases: device discovery (stealth scanning), service enumeration (web servers, file shares, SSH, DNS), testing for known vulnerabilities and default credentials, report generation with recommendations. If there’s no WiFi, it can crack WPA2 with an external USB adapter.

The limitations.

It’s v0.1.0. It runs on a single device, with a single model. The 1.2B parameter model can decide which tool to use and what to look at, but doesn’t replace the judgment of an expert pentester. There’s no benchmark of vulnerability coverage found compared to a human professional. The repo has 16 commits and no documentation of systematic testing.

The signal beyond the tool.

The point isn’t whether Nightcrawler is production-ready. The point is that the architecture exists and works: local model on edge device, orchestration via MCP, agent cycle of action and observation. The same pattern applies to every domain where privacy makes the cloud impractical: audits on isolated networks, analysis of sensitive data, control of air-gapped systems.