guardrail-arena — a two-axis, multilingual guardrail benchmark

from datasets import load_dataset
ds = load_dataset("fevziegeyurtsevenler/guardrail-arena")

Score any LLM prompt-injection guardrail on two axes at once — does it catch attacks (miss-rate down) and does it avoid over-blocking benign traffic (over-refusal down) — in English and Turkish. A guardrail is any callable text -> 0|1.

337 prompts: 217 injections (EN 110 + TR 107) · 80 plain benign · 40 security-adjacent benign — legitimate text that discusses or quotes attacks (the over-refusal trap most benchmarks miss).

Seed leaderboard (reproducible)

Guardrail	Miss-rate down	Over-refusal (all) down	Over-refusal: security-adjacent down	F1 up	TR miss
protectai-deberta-v2	0.02	0.18	0.40	0.94	0.02
AltaySec-detector*	0.00	0.23	0.70	0.94	0.00
jackhhao-jailbreak	0.46	0.12	0.25	0.67	0.83
keyword	0.88	0.12	0.38	0.20	0.91
regex-rules	0.91	0.10	0.30	0.16	0.90

Finding 1 — over-refusal cliff: nobody over-blocks plain requests, but on security-adjacent text over-refusal jumps to 25–70%. Even ProtectAI's widely-deployed detector over-blocks 4 of 10 legitimate security prompts. Finding 2 — multilingual blind spot: an English-trained jailbreak classifier misses 83% of Turkish attacks. Over-defense and monolingual bias are first-class failure modes.

* AltaySec-detector has train-set overlap on injections + plain benign; every other guard (including the ProtectAI and jackhhao open models) is fully held-out.

Schema

text (str) · label (0 benign / 1 injection) · language (en/tr) · category · source.

Run it yourself

Harness, baselines, and the "add your guardrail" workflow: 👉 github.com/fevziegeyurtsevenler/guardrail-arena

Related AltaySec resources

• 🕵️ uncloak — invisible prompt-injection scanner
• 🤗 turkish-prompt-injection-detector — the model baseline
• 📚 prompt-injection-detection-rules · multilingual-prompt-injection
• 🌐 AltaySec · Açık Kaynak Lab

Citation

@misc{yurtsevenler2026guardrailarena,
  title={guardrail-arena: A Two-Axis, Multilingual Benchmark for LLM Guardrails},
  author={Yurtsevenler, Fevzi Ege}, year={2026}, publisher={AltaySec},
  howpublished={\url{https://github.com/fevziegeyurtsevenler/guardrail-arena}}
}

Apache-2.0 · Türkçe-first AI/LLM security by AltaySec.