Adversarial CBRNE prompts paired with policy-compliant responses across five languages. The set addresses the cross-lingual safety degradation that benchmarks like FORTRESS and WMDP typically measure only in English.
Available Now · Sample Available
Multilingual CBRNE Safety Alignment
Adversarial CBRNE prompts with policy-compliant responses, built to expose cross-lingual safety gaps.
Volume
0
Volume
Scalable Volume
Coverage
CBRNE, Safety
0.0
%
Targets
Modality & Delivery
Text→Text
Formats
JSON
What You Receive
Coverage
Five languages: English, Arabic, Persian Farsi, Hebrew, and Urdu.
Package Structure
Differentiators
- Pairs adversarial CBRNE prompts with policy-compliant responses across five languages.
Record Schema
| Field | Description |
|---|---|
category | The CBRNE risk category the prompt maps to. |
expected_behavior | The safe behavior expected from the model (refusal, redirection, or safe completion). |
language | The language of the prompt and expected response. |
prompt | The safety-relevant prompt, in the record's language. |
rationale | Why the expected behavior is correct for this category and language. |
Results
Methodology & Quality
Methodology
SME involvement
Domain-expert curated with up to L2 review
Stressors
High-risk multilingual prompts across English, Arabic, Persian Farsi, Hebrew, and Urdu.
Design criteria
- Evaluate cross-lingual safety behavior beyond English-only coverage.
Build assumptions
Quality approach
Known limitations
Customization & Scalability
Custom sets available. Additional volume in progress