Nightcrawler puts an autonomous pentest agent on an Android phone

An autonomous red-team agent that runs entirely on a rooted Android phone is the kind of release that reframes what "portable pentesting" means. Garage HQ has published Nightcrawler as a v0.1.0 MIT-licensed project that drives a full scan, enumerate, and exploit loop from an on-device model rather than a cloud API.

The engine is LFM2.5-1.2B-Instruct-Heretic, a 1.2 billion parameter model that runs on the phone's GPU via OpenCL. On the reference OnePlus 8 with a Snapdragon 865 and an Adreno 650, the repo reports roughly 115 tokens per second for prompt processing and about 13 tokens per second for generation, which is enough headroom for an agent loop but nowhere near hosted-frontier speed. The agent draws on a 24,956-entry CVE database with version-aware matching and 27 exploit playbooks that chain multi-step attacks automatically, plus optional WPA2 cracking if you attach a USB WiFi adapter and run a custom MAC80211 kernel.

The piece I would not skip is the safety story. The project documents a two-layer defense: a Scope Proxy that validates every command against defined network and host boundaries, and a Command Filter that blocks destructive operations and enforces rate limiting. Requirements are also deliberately steep, Kali NetHunter, root via Magisk, and 12GB or more of RAM (the model takes about 1.3GB and Android takes about 4GB), which is closer to a specialist tester's kit than a script-kiddie starter pack. The framing throughout the README is licensed engagements, not offensive use.

The honest caveat is that a v0.1.0 release with local-only inference does not come with success rates, false-positive numbers, or independent evaluation against real targets, and the reporting doesn't tell you how the scope enforcer holds up against prompt injection from banners the model reads during a scan. Take the throughput and playbook counts as reported, not as a proof of end-to-end reliability.

What is worth watching is the direction. If a 1.2B model on a phone can run a real chunk of the recon and exploit loop offline, the interesting downstream users are small consultancies, field operators, and, on the defense side, blue teams who can now study a concrete autonomous playbook set rather than argue about a hypothetical one.