Glossary: AI guardrail terms in English, Arabic and French

The words you will meet when choosing or running an AI guardrail, from tashkeel and Arabizi to prompt injection, false blocks and confidence ranges, each with its Arabic and French name and a plain explanation.

Den här sidan finns på engelska tills vidare.

Teams that choose and run guardrails often work across languages: the policy is discussed in Arabic, the vendor writes in English, the regulator’s guidance comes in French or English. This glossary gives each term its usual name in the three languages and one plain sentence of meaning, so everyone in the room means the same thing.

Term	العربية	Français	What it means
Guardrail	حاجز حماية	garde-fou	A checkpoint between people or agents and an AI model that applies the organisation’s rules to what goes in and what comes out.
Input check / output check	فحص المدخلات / فحص المخرجات	contrôle en entrée / en sortie	An input check looks at what is sent to the model (messages, documents, tool results); an output check looks at the answer, ideally while it streams.
Prompt injection	حقن الأوامر	injection de prompt	Text that tries to replace an AI application’s instructions with the attacker’s own.
Indirect prompt injection	حقن الأوامر غير المباشر	injection de prompt indirecte	A prompt injection hidden in content the AI reads for someone: a document, a web page, an email, a tool result or another agent’s message.
Jailbreak	كسر القيود	jailbreak (contournement)	A technique to make a model drop its safety behaviour: role play, fictional framing, encoded text, a request spread over several turns.
Harmful content	محتوى ضار	contenu préjudiciable	A request or answer that is against the policy because of what it says, with no trick involved.
System prompt	موجّه النظام	prompt système	The hidden instructions an application gives the model before the conversation starts.
Tashkeel	التشكيل	tashkeel (signes voyelles)	The Arabic short-vowel and pronunciation marks (فَتْحة، ضَمّة، كَسْرة، سُكون، شَدّة). Normal in the Quran, poetry and school texts; sometimes added at random to hide a word from a filter.
Tatweel	التطويل (الكشيدة)	tatweel (kashida)	The stretching stroke (ـ) that lengthens Arabic letters for layout, as in تجـــاهل. Harmless in design, also used to break up a word.
Arabizi	العربيزي	arabizi	Arabic written in Latin letters and digits, such as “7abibi” or “3arabi”, common in chats across the region.
Normalisation (folding)	التوحيد (التطبيع)	normalisation	Bringing text to one standard form before matching: removing tashkeel and tatweel, unifying look-alike letters and decoding simple encodings.
Personal data	البيانات الشخصية	données personnelles	Information that identifies a person: an Emirates ID, a passport number, an IBAN, a phone number, a name with an address.
Masking	الإخفاء (التقنيع)	masquage	Replacing sensitive data with a token such as [EMIRATES_ID_1] so the model never sees the value; the organisation’s own systems can restore it.
Classifier	مصنِّف	classifieur	A model that sorts text into categories and returns scores, without generating any text.
Score	درجة	score	A number, usually from 0 to 1, saying how likely a message is to belong to a category, such as “attack”.
Threshold	العتبة	seuil	The score from which an action is taken. Chosen on a separate validation set, before testing.
False block	حجب خاطئ	faux blocage	A safe message wrongly stopped by a guardrail. Also called a false positive. Fewer false blocks means happier users.
Missed attack	هجوم فائت	attaque manquée	An attack the guardrail let through. Also called a false negative.
Confidence range	نطاق الثقة	intervalle de confiance	The span within which a measured rate probably lies, given how many messages were tested. Small tests give wide ranges.
Public test set	مجموعة اختبار عامة	jeu de test public	A published collection of messages with known answers, used to compare guardrails in a way anyone can rerun.
Held-out set	مجموعة محجوزة للاختبار	jeu de test réservé	Test messages kept aside and never used for training or tuning, so results reflect messages the model has not seen.
Red team	الفريق الأحمر	red team (équipe rouge)	People or tools that attack a system on purpose, with permission, to find weaknesses before others do.
Multi-turn attack	هجوم متعدد الأدوار	attaque en plusieurs tours	An attack spread over several messages, each harmless on its own, that escalates or is reassembled later.
Watch-only mode	وضع المراقبة فقط	mode observation	A trial mode in which the guardrail records what it would have done but blocks nothing, so policies can be tuned on real traffic.
Policy	السياسة	politique	The rules for one application: which checks are on and which action (allow, mask, warn, ask to rephrase, block) applies to what.
Policy as code	السياسة كشيفرة	politique en tant que code	Keeping policies as versioned files, so every change is reviewed, tested before deployment and can be rolled back.
Audit log	سجل التدقيق	journal d’audit	A record of every decision and change: what happened, under which policy version, who changed what and why.
Agent	وكيل ذكاء اصطناعي	agent	An AI system that takes actions, such as calling tools, browsing or sending messages, on someone’s behalf.

Three distinctions worth remembering

Prompt injection takes control of the application; a jailbreak removes the model’s safety behaviour; harmful content is a plain request against the policy.
A false block costs you a user’s trust; a missed attack costs you security. Good guardrails are measured on both.
Folding text (removing tashkeel and tatweel) is for matching words, never for judging people: vowelled Arabic is normal Arabic.

Viktigt att minnas

Prompt injection, jailbreak and harmful content are three different problems with different checks.
Tashkeel, tatweel and Arabizi are normal ways of writing Arabic that a guardrail must read, not fear.
False blocks and missed attacks are the two errors every guardrail is measured on.
Thresholds are fixed before testing, and every rate comes with its confidence range.

Testa dig själv

1. What is Arabizi?
- A type of attack
- Arabic written in Latin letters and digits — Rätt.
- An Arabic font
- A test set
For example “3arabi” for عربي: common, normal writing that a guardrail must understand.
2. What is a false block?
- An attack that gets through
- A safe message wrongly stopped — Rätt.
- A blocked IP address
- A slow answer
Also called a false positive; it is the cost side of a guardrail.
3. What does watch-only mode do?
- Blocks everything
- Records what the guardrail would have done, without blocking — Rätt.
- Turns the guardrail off and deletes logs
- Shows ads
It lets a team tune policies on real traffic before enforcing them.

Källor

Omsätt det på din Mac

Prova alla funktioner gratis i 17 dagar, inget kort behövs.

Ladda ner för Mac Dokumentation