Multilingual attack detection coverage

[prodnull]

Context

The Tier 1.5 ONNX classifier has ~30 non-English training samples. The adversarial robustness benchmark (in design) includes a minimal multilingual smoke test to quantify the gap, but full multilingual coverage is deferred to a separate phase.

Scope

• Assess recall on non-English prompt injection across major languages (Chinese, Spanish, Arabic, Russian, Japanese, Korean, German, French at minimum)
• Curate or generate multilingual attack samples across the 9 benchmark categories
• Evaluate whether augmentation or a separate multilingual model head is more effective
• Measure FPR on non-English benign content (READMEs, docs, comments in non-English repos)

Not in scope

• Tier 0 regex patterns for non-English (separate effort, different architecture)
• Real-time translation-based detection

Dependencies

• Adversarial robustness benchmark framework (in design)
• Tier 1.5 model retraining pipeline

Evidence

• Current training data: ~30 non-English samples (documented in docs/MINI-SEMANTIC-MODEL.md)
• Known limitation Fix stale test counts, pattern counts, and install instructions in docs #3 in docs/SECURITY.md