Across the Gulf, banks, telecoms, hospitals and government services are giving their customers AI assistants that answer in seconds, at any hour. It is one of the most exciting shifts in how people here are served. We want those assistants to be trusted by the companies that run them and loved by the people who use them. That is why we, the HisnLabs team behind FireAI, are building Guardrino: an AI firewall designed from the first line for Arabic and English, and for the data rules of the region.
This note explains what we saw, what we are building, what we will not claim until we have measured it, and how a company can help shape it.
What we saw: people here write in more than one language at once
A customer in Dubai may open with Gulf dialect, switch to English for a product name, and finish in Arabizi, Arabic written with Latin letters and numerals, where 3 stands for ع and 7 for ح, the style people use every day in chat and text messages [1]. A student in Riyadh may paste a fully vowelled verse for a school assistant to explain; most Arabic is written without those marks, but they are standard in the Quran, poetry and teaching material [2]. This is not an edge case. It is how the region talks.
The public tools for testing guardrails were not built around this. GuardBench, one of the broadest collections, gathers 40 safety datasets, and its authors added new moderation data in German, French, Italian and Spanish [3]. CyberSecEval’s prompt-injection tests come as an English file and a machine-translated multilingual file [4]. Both are valuable, and we use them. Neither covers Gulf dialect or Arabizi as people actually write them. A guardrail that is only measured on English cannot promise much to an Arabic-speaking customer, in either direction: it may miss a manipulation attempt written in dialect, or refuse a perfectly ordinary question because it looks unfamiliar.
What we saw: data rules that ask where the data goes
The second reason is regulation. The UAE’s Federal Decree-Law No. 45 of 2021, in force since 2 January 2022, defines the controls for processing personal data and “sets out the requirements for the cross-border transfer and sharing of personal data for processing purposes” [5]. Saudi Arabia issued its Personal Data Protection Law on 16 September 2021, administered through SDAIA [6]. For a company that sends customer messages to an AI model hosted abroad, these rules turn a technical choice into a legal one.
So a guardrail for this region has to do two things at once: keep personal data such as Emirates IDs, Saudi IDs and IBANs away from the model provider, and run where the company needs it to run.
What Guardrino is designed to do
Guardrino is planned as a gateway between a company’s staff, its AI agents and any LLM. Everything in this list is design, not a released feature.
- Read Arabic as it is written: Modern Standard Arabic, Gulf and Egyptian dialects, Arabizi and mixed Arabic and English, with diacritics, stretched letters and look-alike characters normalised before any check.
- Check every message in layers: fast checks for the clear cases, a small language-aware classifier for manipulation, and a deeper review only for the few messages that remain unclear.
- Mask Gulf personal data, checked by its structure and check digits, before the model sees it, and put the real value back in the answer.
- Let your policy decide among five outcomes, allow, mask, warn, ask to rephrase or block, per app and per team, so normal users are not turned away.
- Inspect what agents read, such as documents and tool results, because OWASP counts injection through “external sources, such as websites or files” as part of the same risk [7].
- Run in the UAE as a hosted gateway, or entirely inside your own network as an offline install.
The details of the layers are in our plain guide to AI guardrails, and the way Guardrino is designed to keep legitimate users happy is in guardrails that don’t annoy your users.
What we will not claim until we have measured it
Guardrino is in development. There is no public version, no price, no release date and no customer yet, and we do not publish performance figures for something that has not been released. The Guardrino page has a section for measured results that stays empty by design. It fills in only when a release has been tested on public test sets and on a held-out set of our own in Arabic and English, with every number reported next to the number of messages behind it and its confidence range. We wrote down that method in public: how to test an AI guardrail with public test sets. You can hold us to it, and you can use it on anyone else.
We think this is the confident way to build a security product. A buyer who can check the method does not need to take our word for the result.
How a company can join the pilot
We are looking for a small number of companies in the UAE and the Gulf that run, or plan to run, AI apps or agents in Arabic and English. The pilot is designed to be low-risk for you.
- A short call about your AI apps, your languages and your rules.
- A watch-only trial on your own traffic, in our UAE cloud or inside your network. Guardrino records what it would have done and blocks nothing.
- A report of what Guardrino would have masked, warned about or blocked, which we review with you.
- Your feedback shapes what ships first, including which deployment option comes first.
If that sounds like your team, visit the Guardrino page to see the planned design and the browser-only demo, or go straight to request early access. We would be glad to hear how your customers write.
How Guardrino and HisnLabs fit in
Guardrino is in development, and the pilot programme is the only way to use it early. The simulation below shows the kind of messages it is being designed for: Arabic, Arabizi and the personal data people in the Gulf type every day.