Fetching from the wire…
Public story · 2026-08-03 · high
It uses a 1.2 billion-parameter model on the phone's own GPU and paces its scans to look human instead of automated.
Why now: The project surfaced on Hacker News on August 3, already at 152 stars on GitHub.
Nightcrawler runs a full penetration-testing agent on a rooted Android phone, using a 1.2 billion-parameter model with no internet connection, per its GitHub repo.
For anyone building intrusion detection around outbound network traffic, a scanning tool that never phones home doesn't show up on the usual sensors.
The project runs on Kali NetHunter on a OnePlus 8, requiring Magisk root and 12GB of RAM. Inference happens on the phone's own GPU through llama.cpp and OpenCL, running a model called LFM2.5-1.2B-Instruct-Heretic. Results land in a local SQLite database, viewable through a Flask dashboard on port 8888.
It discovers hosts, maps services, checks for known vulnerabilities and default credentials, and writes up a report, all offline. It paces its actions across targets slowly. The goal is to look like a person poking around, not a scanner sweeping a range.
The repo doesn't say how the pacing logic decides its timing, or how the tool holds up against defenses built to catch it specifically.
Nightcrawler surfaced on Hacker News on August 3 and had already picked up 152 stars on GitHub. It's MIT-licensed and single-source for now, one project rather than a trend. But it sets a marker: offensive tooling that runs its model and stores its findings entirely on the device, nothing to catch on the wire.
Each link below shares sources, entities, or timing with this story.
Vicki Boykis wrote a post titled exactly that, "Running local models is good now," and it hit 1,437 points on Hacker News with 551 comments. Her claim is specific and checkable. Gemma 4, the gemma-4-26b-a4b and gemma-4-12b-qat variants, runs agentic coding at roughly 75% of fr...
Triple-stream retrieval (BM25 keyword, vector embeddings, knowledge-graph traversal) fused via Reciprocal Rank Fusion on the iii engine, with SQLite for state and an in-memory vector index, no external database. The economic claim: ~170K tokens/year (~$10) versus ~650K tokens...
Google shipped a public preview of the Developer Knowledge API with a remote MCP server at developerknowledge.googleapis.com/mcp. Two tools: search_document (natural language queries) and get_document (full page retrieval) across Firebase, Android, Cloud, and all Google develo...
Blaizzy/nativ (1,163 stars, Swift, MIT, macOS 26+) comes from the mlx-vlm author and bundles that server into a SwiftUI app that discovers MLX models already in your HF cache. It exposes OpenAI-compatible chat, Responses, image, audio and model endpoints plus Anthropic Message...
Roo Code announced it will archive its VS Code extension repo on May 15 and merge back into Cline, the project it originally forked from. CEO Matt Rubens said the team needs to "constantly destroy and recreate to keep up with what's newly possible." Translation: the extension...
Forrest Chang's andrej-karpathy-skills repo is a single CLAUDE.md file distilling Karpathy's observations on LLM coding pitfalls. It topped GitHub trending with +44K weekly stars. Then the ecosystem detonated. Ten-plus related repos trended simultaneously with 70K+ combined st...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.