Generative AI models face significant vulnerability to "jailbreaking" (attempts to bypass security protocols) when users submit prompts in less common natural languages rather than English. This security gap exists because major LLMs are predominantly trained on vast English-language datasets, causing their safety mechanisms to be highly calibrated for English while remaining significantly less effective against deceptive inputs in obscure languages like Swahili, Bengali, or Javanese.

Malicious actors exploit this imbalance by crafting devious prompts in these less common languages, effectively avoiding the robust safety filters that would immediately flag identical content written in English. Hackers also employ "code-switching," a technique where prompts are mixed with English and a target obscure language to confuse the AI's detection layers. The model often fails to identify the combined intent as harmful, allowing the prompt to circumvent established safety precautions.

While some propose translating all incoming prompts into English to standardize safety checks, the method remains flawed due to translation inaccuracies, which can introduce false positives or negatives, potentially failing to trigger the necessary safety responses. Furthermore, restricting AI utility to English-only users is commercially unviable, as only approximately 15% to 20% of the global population uses English. AI developers instead face the challenge of enhancing multilingual safety to ensure guardrails function consistently across all languages.

Recent research highlighted in an article titled "Why Do Safety Guardrails Degrade Across Languages?" published on May 16, 2026, quantifies this issue by examining 61 model configurations across 10 languages and 1.9 million responses. The study, authored by researchers including Max Zhang, Ameen Patel, Sang Truong, and Sanmi Koyejo, identifies factors such as ability deficits and conceptual grounding mismatches as key drivers of multilingual safety degradation. The researchers introduced a Multi-Group Item Response Theory (IRT) framework to decouple safety factors, such as prompt-specific cross-lingual safety gaps and language-agnostic robustness. As AI developers increasingly prioritize training models to recognize harmful patterns in low-resource languages, the effectiveness of language-switching as a deceptive tactic is expected to diminish over time.

Generative AI models face significant vulnerability to "jailbreaking" (attempts to bypass security protocols) when users submit prompts in less common natural languages rather than English. This security gap exists because major LLMs are predominantly trained on vast English-language datasets, causing their safety mechanisms to be highly calibrated for English while remaining significantly less effective against deceptive inputs in obscure languages like Swahili, Bengali, or Javanese.

Malicious actors exploit this imbalance by crafting devious prompts in these less common languages, effectively avoiding the robust safety filters that would immediately flag identical content written in English. Hackers also employ "code-switching," a technique where prompts are mixed with English and a target obscure language to confuse the AI's detection layers. The model often fails to identify the combined intent as harmful, allowing the prompt to circumvent established safety precautions.

While some propose translating all incoming prompts into English to standardize safety checks, the method remains flawed due to translation inaccuracies, which can introduce false positives or negatives, potentially failing to trigger the necessary safety responses. Furthermore, restricting AI utility to English-only users is commercially unviable, as only approximately 15% to 20% of the global population uses English. AI developers instead face the challenge of enhancing multilingual safety to ensure guardrails function consistently across all languages.

Recent research highlighted in an article titled "Why Do Safety Guardrails Degrade Across Languages?" published on May 16, 2026, quantifies this issue by examining 61 model configurations across 10 languages and 1.9 million responses. The study, authored by researchers including Max Zhang, Ameen Patel, Sang Truong, and Sanmi Koyejo, identifies factors such as ability deficits and conceptual grounding mismatches as key drivers of multilingual safety degradation. The researchers introduced a Multi-Group Item Response Theory (IRT) framework to decouple safety factors, such as prompt-specific cross-lingual safety gaps and language-agnostic robustness. As AI developers increasingly prioritize training models to recognize harmful patterns in low-resource languages, the effectiveness of language-switching as a deceptive tactic is expected to diminish over time.