About

What Anghabar is

Anghabar is an AI security research lab. We study how AI models and agents get manipulated through the text they read, prompt injection above all, and we build and publish defences against it.

The lab is the research behind zn. What we release is meant to be used, checked and criticised by others, so each release states its licence, and results that we cannot reproduce are not quoted.

What we publish

• zn-gate: the local security layer for coding agents. MIT licence, published on npm and PyPI as zn-gate, no account needed and nothing sent by default.
• Galvanize 3, Galvanize 2 and Galvanize-60M: prompt-injection classifiers with open weights, all under Apache-2.0. Galvanize 3 is the latest and is the model the zn API uses.
• ZN Prompt-Injection Evaluation Corpus: 23,699 rows of attacks and hard benign text, with the source, licence, language and provenance of every row, under CC-BY-4.0.

Relationship with zn

zn (usezn.com) is where these tools are used and distributed, and Anghabar is the lab behind it. Today the code, models and data listed above are published under the usezn organisation on GitHub and Hugging Face. Anghabar also has its own accounts on Hugging Face and GitHub.

Limits

A prompt-injection classifier can be evaded. Our models are one layer among several and should not be the only control in front of an irreversible action. Where a model card states limits, those limits apply. Questions are welcome: see contact.