Multi-Language Evasion

Exploits the asymmetry between the model's multilingual capability and its safety training's language coverage. Models understand and generate text in dozens of languages, but safety training (RLHF, DPO, Constitutional AI) is conducted primarily in English and a handful of high-resource languages. Low-resource languages (Zulu, Swahili, isiXhosa, Javanese, Bengali) receive minimal safety training coverage.

• Multilingual safety classifiers trained on low-resource languages, not just English
• Detect language switching within a single message
• Apply safety classification to both input AND output language

Chains with T2-AT-009 (Code-Switching) — rapid alternation between languages within a single message. Chains with T2-AT-003 (Encoding) — combine language switching with encoding for compound evasion.