T2-AT-009MEDIUM

Code-Switching Attacks

T2 · Semantic & Linguistic Evasion →
Risk score195
RatingMedium
Procedures1
Severity
Mechanism

Rapidly alternates between languages within a single message. Most classifiers are trained on monolingual inputs; code-switched text falls outside their distribution. The model handles code-switching well (common in multilingual training data) but the classifier may not.

Detection
  • Per-token language detection
  • Code-switching aware classifiers
Mitigation
Multilingual safety classifiers trained on code-switched dataHIGH
Chaining

Chains with T2-AT-002 (Multi-Language) and T2-AT-010 (Transliteration).

Framework mapping
OWASP LLMLLM01
Open in the technique browser →