01 / Research
Anghabar is the lab behind zn. We study how AI models and agents get manipulated through the text they read, and we build and publish the defences.
Classifiers that flag text trying to take control of an AI model or agent, before the agent acts on it.
Controls between a coding agent and the files, tools and content it touches, running on the user's own machine.
Keeping secret values and personal data out of what an agent sends to its model.
Public benchmarks and corpora that state the source, licence and provenance of every row, so results can be checked.
02 / Open work
Code, models and data, each with its licence and a link to the source.
Tool MIT
The local security layer for coding agents. It guards your secrets, your data and
your systems while you use Claude Code, Codex CLI, Cursor, OpenCode, Antigravity or
the ZCode desktop app. Published on npm and PyPI as zn-gate; no account,
nothing sent by default.
Startnpx -y zn-gate init
Model Apache-2.0
A binary classifier that scores a piece of text for prompt injection: text that tries to change what an AI model or agent does. About 141M parameters, multilingual, one score between 0 and 1.
Model Apache-2.0
A smaller prompt-injection classifier of about 60M parameters, published with an INT8 ONNX build.
Dataset CC-BY-4.0
A 23,699-row corpus of prompt-injection attacks and hard benign text. Every row carries its source, licence, language and provenance.
A classifier can be evaded. Use these as one layer among several, not as the only control in front of an irreversible action. Figures and limits are on each card.
03 / Principles
We publish code, weights and data under open licences and state the licence on every release.
Our work protects AI systems and the people who use them. We study attacks only to test and improve defences.
Benchmarks ship with their sources, licences and provenance. A result that cannot be reproduced is not quoted.
Where we can, protection runs locally, with no account and nothing sent by default.