I built Nightcrawler, an open-source autonomous penetration-testing agent that runs entirely on an Android phone.
The project started with a question: how much of a real pentesting workflow could I run locally on relatively old mobile hardware, without relying on a cloud model or API?
Nightcrawler runs a 1.2B-parameter model locally on the Adreno GPU of a OnePlus 8. The model chooses targets and tools, while a separate scope-enforcement proxy validates every command before execution. The system maintains per-host memory in SQLite, rotates between targets, matches detected versions against a local CVE database, executes multi-step playbooks, and generates a structured report.
A few implementation details that may be interesting:
Local inference runs at roughly 115 prompt tokens/sec and 13 generated tokens/sec.
The small model only produces a usable command around 50% of the time, so much of the engineering is recovery logic, duplicate detection, persistent memory, and deterministic playbooks.
Every command passes through a separate scope and safety layer rather than trusting the model to remain in scope.
The project includes a dry-run mode, so the agent loop can be tested without executing real network commands or owning the phone hardware.
I've had it running on my home network for the past 3 months uninterrupted
The following rant is not against the owner/project - but...
What an irony. I cant publish a attack surface mapping / pentesting tool i wrote which runs fully deterministic and really controlable due to "dual use" legal problems - but llm driven tools hit public space......
What does the 50% look like when it fails? Garbage the parser throws out is easy to handle, but a well formed command aimed at the wrong host gets past the scope check, and you would only catch that reading the report afterward.
This looks interesting. My current phone is an Asus ROG Phone 3, which I think suffices (Snapdragon 865+ plus 12Gb RAM) but isn't rooted. Even on outdated Android 12 this phone covers everything I could possibly need as-is, and I don't intend to replace it. But, this might be the final push to upgrade ROM to something newer. I saw crDroid 16.0 still supports this ROG variant.
Will this work with Nethunter Rootless, or ir Magisk support w/root imperative?
Fantastic project! Out of curiosity, do you foresee any optimization paths (like model quantization) that could lower the memory footprint down the road
14 comments
[ 1.6 ms ] story [ 25.6 ms ] threadThe project started with a question: how much of a real pentesting workflow could I run locally on relatively old mobile hardware, without relying on a cloud model or API?
Nightcrawler runs a 1.2B-parameter model locally on the Adreno GPU of a OnePlus 8. The model chooses targets and tools, while a separate scope-enforcement proxy validates every command before execution. The system maintains per-host memory in SQLite, rotates between targets, matches detected versions against a local CVE database, executes multi-step playbooks, and generates a structured report.
A few implementation details that may be interesting:
Local inference runs at roughly 115 prompt tokens/sec and 13 generated tokens/sec. The small model only produces a usable command around 50% of the time, so much of the engineering is recovery logic, duplicate detection, persistent memory, and deterministic playbooks. Every command passes through a separate scope and safety layer rather than trusting the model to remain in scope. The project includes a dry-run mode, so the agent loop can be tested without executing real network commands or owning the phone hardware. I've had it running on my home network for the past 3 months uninterrupted
What an irony. I cant publish a attack surface mapping / pentesting tool i wrote which runs fully deterministic and really controlable due to "dual use" legal problems - but llm driven tools hit public space......
sorry for the rant....
Will this work with Nethunter Rootless, or ir Magisk support w/root imperative?