Multilingual safety-eval for a frontier AI lab

The lab's flagship model passed English safety benchmarks but had no credible way to measure — let alone catch — culturally-specific harms in Indic languages. Caste slurs, communal dog-whistles, regional abuse and code-mixed toxicity slipped past guardrails trained on translated Western taxonomies.

They needed an evaluation set that reflected how harm is actually expressed across ten Indic languages, with measured agreement they could defend to their own safety and policy teams.

We built a native-speaker red-team and safety-eval program rather than translating an existing benchmark:

• Recruited vetted native-speaker linguists across ten languages under written contributor agreements.
• Co-designed a culturally-grounded harm taxonomy (11 categories) with the lab's policy team, piloted before scale.
• Authored adversarial prompts from scratch with escalation rounds and a cultural-plausibility review — no scraped forum data.
• Ran two-pass QA with adjudication; reported Krippendorff's α per category.

The lab received a documented evaluation set with inter-annotator agreement of 0.86 across 11 harm categories and a post-adjudication QA pass rate above 98%. It exposed material gaps in Hindi and Tamil guardrail coverage that English benchmarks had hidden, and became a recurring quarterly refresh as new harm patterns emerged.

Representative outcome from an anonymized engagement; metrics are illustrative of our standard quality bar.