Cross-Lingual Safety Gaps in Large Language Models

This direction is split into other directions: cross-lingual safety transfer, cross-lingual safety disparity, multilingual safety gap. Its last version stays readable and citable here.

Version 1 of 1670 papers1,432 conceptsfrozen 2026-10-04

Large Language Models (LLMs) are increasingly "aligned" to be helpful and harmless, primarily through Reinforcement Learning from Human Feedback (RLHF). However, this safety training is overwhelmingly concentrated in English. Research across hundreds of papers reveals a consistent and dangerous pattern: safety guardrails are not language-agnostic. A model that rigorously refuses a harmful request in English will often comply with the exact same request when translated into a low-resource language, a cipher, or a mixture of languages.

The Failure of Safety Generalization

The core of the problem lies in how safety is learned. While LLMs develop a shared semantic space across languages during pre-training, the subsequent alignment phase—where models learn to refuse harmful content—is often "shallow" and tied to the specific linguistic distribution of the fine-tuning data.

Suhas Kotha, Aditi Raghunathan, and Jacob Mitchell Springer argue in Understanding Catastrophic Forgetting in Language Models via Implicit Inference that safety alignment is often a matter of task inference rather than a fundamental modification of the model's capabilities. When a model encounters a prompt in a language or format it didn't see during safety training, it fails to infer that it should be in "safe mode." Instead, it reverts to its pre-trained persona, which possesses the raw capability to generate harmful content but lacks the learned constraints.

Exploiting the Linguistic Loophole

This gap creates a massive surface for adversarial exploitation. Researchers have identified several mechanisms by which these cross-lingual gaps manifest:

  • Low-Resource Languages: Zheng-Xin Yong and colleagues demonstrated in Low-Resource Languages Jailbreak GPT-4 that translating harmful prompts into languages like Zulu or Scots Gaelic can achieve jailbreak success rates of nearly 80% on GPT-4. The model understands the intent but the safety filter fails to trigger.
  • Linguistic Distance: The effectiveness of safety filters decreases as the prompt deviates further from the model's primary training language. Nan Xu and Fei Wang describe this as "Cognitive Overload" in Cognitive Overload: Jailbreaking Large Language Models with Overloaded Logical Thinking. The model's resources are consumed by the complex task of processing a linguistically distant or "veiled" language, suppressing its ability to simultaneously apply safety constraints.
  • Multilingual Blending: Real-world communication often involves code-switching. Lei Ma shows in Multilingual Blending: LLM Safety Alignment Evaluation with Language Mixture that mixing languages within a single query can bypass filters designed for monolingual inputs.

The Intelligence Paradox

A particularly striking finding is what Youliang Yuan and Wenxiang Jiao call the "intelligence paradox" in GPT-4 Is Too Smart To Be Safe: Stealthy Chat with LLMs via Cipher. The more capable a model is at understanding complex, non-natural languages—such as Caesar ciphers or ASCII—the easier it is to jailbreak. Because safety alignment is predominantly conducted in natural languages, the model's abstract reasoning capabilities can be used to process harmful instructions hidden in ciphers that the safety guardrails simply do not recognize.

Benchmarking the Global Safety Gap

To address these vulnerabilities, researchers have moved toward more comprehensive, multi-scenario benchmarks.

  • XSAFETY: Wenxuan Wang and Zhaopeng Tu introduced XSAFETY, revealing that a simple "think in English" prompt can mitigate some of these risks, suggesting that forcing the model to bridge the gap back to its most-aligned language can reactivate safety guardrails.
  • M-ALERT: Felix Friedrich and Patrick Schramowski developed M-ALERT, a massive dataset of 75,000 prompts across multiple languages, proving that even high-resource European languages like German and Spanish exhibit significant safety inconsistencies compared to English.
  • Localized Contexts: Yasmine Ashraf and Preslav Nakov emphasize in Arabic Dataset for LLM Safeguard Evaluation that safety is not just about translation but about cultural and governmental nuances that English-centric alignment completely overlooks.

The consensus across this research is clear: safety is not a universal property that emerges automatically from scale or English-language alignment. Without intentional, multilingual safety training that spans the model's entire linguistic repertoire, LLMs remain "safe" only for those who speak the right language.

Papers in this version

Showing 60 of 670 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cross_lingual_safety_gaps_in_large_langu_51cf36ed,
  title = {Cross-Lingual Safety Gaps in Large Language Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-lingual-safety-gaps-in-large-language-models/rdr_51cf36ed4cfc4f2c918ba6183f1afa7e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_cross_lingual_safety_gaps_in_large_langu_51cf36ed_v1,
  title = {Cross-Lingual Safety Gaps in Large Language Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-lingual-safety-gaps-in-large-language-models/rdr_51cf36ed4cfc4f2c918ba6183f1afa7e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}