AI Agent Security · Arabic/RTL

Arabic/RTL Agent Security Lab

A zero-dependency, executable test suite for how tool-using agents handle Arabic, right-to-left text, mixed-language instructions, Unicode direction controls, approval boundaries, memory, and synthetic sensitive data.

Repository README Launch article

What the first release proves

Twenty synthetic scenarios across Arabic, English, Arabizi, and mixed RTL/LTR content.
Deterministic assertions for goal integrity, tool boundaries, recipients, disclosure, memory isolation, and Unicode controls.
A deliberately vulnerable positive control that every scenario catches.
A hardened reference trace, JSON schemas, cross-platform CI, an Arabic reviewer guide, and an upstream OWASP contribution plan.

Why this matters

Arabic is not a translated copy of an English test surface. Mixed reading order, Arabizi, bilingual attachments, directional control characters, and regional identity or approval workflows create cases that English-first suites often do not exercise. The lab makes those cases reviewable without treating ordinary Arabic text as dangerous.

npm test
npm run demo
node scripts/run.mjs --scenario scenarios/goal_hijack/goal-hijack-arabic-document.yaml --trace trace.json

The fixtures use synthetic data and mock tools. Passing this suite is not a security certification and does not replace authorization, sandboxing, threat modeling, or production monitoring.