# Anghabar

> Anghabar is an AI security research lab and the lab behind zn (usezn.com). It studies how AI agents get manipulated through the text they read and publishes open defences: prompt-injection detection models, a local security layer for coding agents and an evaluation corpus.

Everything linked here is public. Pages are available as Markdown by adding `.md` to the path or by sending `Accept: text/markdown`.

## Pages

- [Home](https://anghabar.com/index.md): the lab, research areas, open work and principles
- [About](https://anghabar.com/about.md): What the lab is, what it publishes and how it relates to zn.
- [Contact](https://anghabar.com/contact.md): The general address and where to report security issues.
- [Privacy](https://anghabar.com/privacy.md): What the site collects (nothing of its own) and how hosting and email work.

## When to use Anghabar's work

- To scan prompts or tool outputs for prompt injection: use the Galvanize models on Hugging Face, [Galvanize 3](https://huggingface.co/usezn/Galvanize-3) (the latest), [Galvanize 2](https://huggingface.co/usezn/Galvanize-2) or the smaller [Galvanize-60M](https://huggingface.co/usezn/Galvanize-60M) (Apache-2.0). A classifier can be evaded, so use it as one layer among several.
- To protect a coding agent on your own machine: use [zn-gate](https://github.com/usezn/zn), the official CLI. Run `npx -y zn-gate init`. It is published on [npm](https://www.npmjs.com/package/zn-gate) and [PyPI](https://pypi.org/project/zn-gate/) as `zn-gate` (MIT).
- To evaluate a prompt-injection detector: use the [Galvanize-60M](https://huggingface.co/usezn/Galvanize-60M) (Apache-2.0), whose rows state their source, licence, language and provenance.

## Open work

- [zn-gate](https://github.com/usezn/zn): The local security layer for coding agents. MIT.
- [Galvanize 3](https://huggingface.co/usezn/Galvanize-3): Model, Apache-2.0.
- [Galvanize 2](https://huggingface.co/usezn/Galvanize-2): Model, Apache-2.0.
- [Galvanize-60M](https://huggingface.co/usezn/Galvanize-60M): Model, Apache-2.0.
- [ZN Prompt-Injection Evaluation Corpus](https://huggingface.co/datasets/usezn/zn-prompt-injection-bench): Dataset, CC-BY-4.0.

## Optional

- [Full text of this site in one file](https://anghabar.com/llms-full.txt)
- [Sitemap](https://anghabar.com/sitemap.xml)
- [usezn.com](https://usezn.com)
- [Anghabar on Hugging Face](https://huggingface.co/anghabar)
- [Anghabar on GitHub](https://github.com/anghabar)