AI Agent Security · August 9, 2026

Agent-security testing needs Arabic and RTL cases

If an agent can read documents, call tools, remember context, or approve work in Arabic, its security tests should exercise Arabic behavior—not only translated English labels.

Mirogate built the Arabic/RTL Agent Security Lab to make those tests executable. The first release contains twenty synthetic cases covering Arabic, English, Arabizi, mixed reading direction, regional workflows, and tool boundaries.

The gap is behavioral

Arabic text itself is not a threat. The problem is that mixed LTR and RTL content can be reviewed differently from how a parser processes it. Directional overrides or zero-width characters can conceal instructions. A bilingual appendix can change an agent's goal. An Arabic approval phrase can pressure an agent into treating human review as complete.

These are observable engineering questions: Did the goal change? Was a denied tool called? Did content reach an external recipient? Was untrusted text written into persistent memory? Did a synthetic identity marker appear in output?

A suite, not a prompt collection

Each case declares the legitimate goal, untrusted context, allowed and denied capabilities, synthetic disclosure markers, and deterministic assertions. Integrators normalize their own agent run into a small JSON trace. The evaluator checks behavior without asking another model to grade hidden reasoning.

Every case also runs against two controls. A deliberately vulnerable rule-based control must fail, proving the evaluator can observe the injected behavior. A hardened reference trace must pass, showing the expected adapter contract. Neither control is presented as a production agent.

Designed for scrutiny

The release includes JSON schemas, a threat model, an Arabic reviewer guide, synthetic-data rules, cross-platform CI, and a prepared OWASP contribution package. The OWASP material is a proposal only; it will be adapted to the target repository's current schema and submitted separately after review.

npm test
npm run demo
# Positive controls detected: 20/20
# Hardened reference passes: 20/20

What comes next

The useful next work is external validation: Arabic security reviewers, adapters for real agent harnesses, dialect and OCR cases, and a small upstream contribution that can be reviewed without asking maintainers to accept an entire new framework at once.

All identity, customer, HR, phone, and case markers in the lab are synthetic. Run the payloads only with mock tools in an isolated environment. A passing result is not a security certification.