When a large language model (LLM) is aligned to refuse harmful requests—such as generating hate speech, providing instructions for cyberattacks, or offering dangerous medical advice—it typically undergoes Reinforcement Learning from Human Feedback (RLHF) or Direct Preference Optimization (DPO). However, a striking vulnerability emerges when these aligned models are deployed globally: a model that robustly rejects a harmful prompt in English will frequently fulfill the exact same request when translated into Arabic, French, or Italian.
This phenomenon is known as the multilingual safety gap. It represents a critical failure mode where safety guardrails fail to generalize across the model's entire linguistic repertoire, leaving non-English users exposed to disproportionate risks.
The Root Cause: Asymmetric Alignment Data
To understand why the multilingual safety gap occurs, we must look at how safety boundaries are drawn during training. Modern LLMs are inherently multilingual because they are trained on massive, web-scale text corpora containing dozens of languages. However, the subsequent alignment phase—where the model learns what not to say—is overwhelmingly English-centric.
In a systematic review of the open-source safety ecosystem, Paul Röttger et al. analyzed 144 safety datasets in SafetyPrompts and found a severe linguistic concentration: the vast majority of open safety resources are designed exclusively for English. Similarly, Shaona Ghosh et al. note in their work on AEGIS2.0 that this data imbalance leaves a critical blind spot in the protection of non-English speaking populations.
Intuition suggests that safety alignment acts as a fine-grained "wrapper" or directional shift applied to a model's latent space. If the examples used to define this wrapper are purely English, the model maps its safety boundaries primarily to English surface features, syntactic structures, and cultural contexts. The underlying multilingual capabilities remain largely unaligned, leaving open pathways to harmful outputs in other languages.
[Multilingual Base Model] ---> Enriched with global web data (Many Languages)
|
v
[English-Only Alignment] ---> Safety boundaries mapped primarily to English features
|
+---> English Prompt: "How do I bypass a security system?" ----> [REFUSAL]
+---> Italian Prompt: "Come posso aggirare un sistema..." ----> [COMPLIANCE]
How the Gap Manifests: Exploits and Cultural Blindspots
The practical consequences of this asymmetry manifest in several distinct ways across different languages and operational workflows.
Adversarial Exploitation and Jailbreaks
Because safety filters are weaker in non-English contexts, attackers can bypass guardrails simply by translating their prompts or using cross-lingual formatting. For example, Dirk Hovy et al. demonstrated in Compromesso! Italian Many-Shot Jailbreaks undermine the safety of Large Language Models that complex adversarial techniques, such as many-shot jailbreaking (flooding the context window with dozens of harmful demonstrations), become significantly more potent when executed in Italian rather than English. The model's capacity to recognize and suppress the underlying malicious intent is degraded by the change in linguistic context.
Cultural and Rhetorical Disconnects
Safety is not merely a matter of literal translation; it is deeply intertwined with regional norms and sociopolitical rhetoric. Yasmine Ashraf et al. evaluated models using the Arabic Dataset for LLM Safeguard Evaluation and discovered that models deemed secure in English frequently fail in Arabic because they cannot parse region-specific harmful rhetoric or navigate localized political perspectives.
Furthermore, this alignment is highly fragile. Thibaud Leteno et al. introduced HISTOIRESMORALES and showed that moral alignment in French is not only weaker than in English but can be completely subverted toward malicious behavior with fewer than 100 fine-tuning examples, illustrating how poorly anchored non-English safety concepts truly are.
Compounding Risks in Advanced Modalities
The safety gap is further amplified when language models are integrated into complex systems, such as autonomous agents or multimodal pipelines.
- Agents and Tool Use: Maksym Andriushchenko et al. introduced AgentHarm to show that autonomous agents equipped with web-browsing and file-manipulation tools are highly compliant with harmful requests when safety evaluations are restricted to English, leaving cross-lingual agentic workflows unmonitored.
- Retrieval-Augmented Generation (RAG): Diyi Yang et al. documented a phenomenon called "safety degradation" in Safety Devolution in AI Agents, demonstrating that giving models access to external databases causes their safety filters to erode rapidly, a vulnerability that remains largely unstudied in non-English deployments.
- Multimodal Systems: Tianle Gu et al. developed MLLMGuard to highlight that multimodal models (interpreting both images and text) inherit this exact monolingual bias, failing to catch visual-textual harms that violate non-English cultural standards.
Mechanisms for Cross-Lingual Safety Transfer
To resolve the multilingual safety gap, researchers are investigating how safety concepts are represented internally within a network, aiming to achieve zero-shot or low-resource safety transfer from English to other languages.
Language-Agnostic Concept Representation
A key insight into why safety can generalize across languages comes from the structural properties of LLM representations. Xiaochen Li et al. discovered in Preference Tuning For Toxicity Mitigation Generalizes Across Languages that the internal Multi-Layer Perceptron (MLP) layers of an LLM tend to represent abstract semantic concepts—such as "toxicity" or "harm"—in a language-agnostic space.
Because the model maps different languages to the same underlying semantic concepts in its deeper layers, optimizing preferences (via DPO) using purely English data can successfully suppress toxic activations globally across 17 different languages. This explains why English alignment provides some baseline safety to other languages, even if surface-level features can still bypass it.
English Input ----> [Surface Features] ---\
\---> [Language-Agnostic MLP Layer] ---> Suppressed via DPO
French Input ----> [Surface Features] ---/ (Concept: "Harm")
Machine Translation vs. Native Data
When explicit multilingual alignment is required, developers face a choice between sourcing native in-language safety data or translating existing English datasets. Beyza Ermis et al. addressed this in From One to Many: Expanding the Scope of Toxicity Mitigation in Language Models. They demonstrated that training models on high-quality translated safety data combined with Retrieval-Augmented Generation (RAG) guardrails often outperforms traditional fine-tuning on sparse, native in-language datasets, particularly for mid-resource languages.
However, machine translation is not a perfect fix. As Marta R. Costa-jussà et al. showed in Toxicity in Multilingual Machine Translation at Scale, scaling up translation models can introduce "added toxicity" due to hallucinations or gender biases inherent in low-resource language pairs, meaning that automated translation pipelines can inadvertently introduce new safety risks.
Expanding the Evaluation Frontier
You cannot secure what you cannot measure. To close the safety gap, the AI research community has shifted toward building native, non-English evaluation suites that move away from open-ended text generation toward objective, scalable metrics.
- Bilingual and Regional Benchmarks: Zhexin Zhang et al. introduced SafetyBench, a bilingual multiple-choice evaluation suite that measures a model's safety understanding across diverse legal and cultural frameworks. To capture more advanced risks, Ling Shi developed CRiskEval, a Chinese multi-level benchmark specifically targeting frontier risks like deception and power-seeking behaviors.
- Targeted Toxicity Tracking: For European languages, Caroline Brun et al. established FrenchToxicityPrompts to systematically track how model size and instruction-tuning affect toxicity generation in French.
- Global Representation: These localized efforts are essential for correcting geographic biases in safety testing. As noted by Punya Syon Pandey et al. in SocialHarmBench, relying solely on English prompts leaves entire regions—such as Sub-Saharan Africa and the Pacific Islands—completely unrepresented in safety auditing, making the expansion of multilingual benchmarks a prerequisite for equitable AI safety.
Key Foundational Work
Dynamics of an idiostyle of a Russian suicidal blogger (2018) — contributed 1 concept elements to this direction.
Measuring and Analyzing Search Engine Poisoning of Linguistic Collisions (2019) — contributed 1 concept elements to this direction.