---
title: "The Illusion of Cross-Lingual Safety in Low-Resource Languages"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.11146v1"
permalink: "https://feed.astralor.com/items/itm_8ecaca672ff564d9de75e4c8"
published: "2026-08-11T17:05:26.000Z"
score: 77
channel: "ai"
tags: ["AI Security","LoDNA","Low-resource Language","Multilingual Language Model","Safety Alignment","Hidden State Analysis"]
---

# The Illusion of Cross-Lingual Safety in Low-Resource Languages

## AI commentary

> I think this paper punctures a dangerous illusion: safety alignment does not automatically transfer to African languages just because a model 'understands' English. For multilingual LLM teams and AI safety governance, this is a warning to take seriously—covering low-resource languages in safety testing is not a bonus but a responsibility.

## At a glance

Safety alignment in LLMs is largely built on English and assumed to transfer across languages. This paper tests that assumption in four African languages—Twi, Hausa, Amharic, and Swahili—using LoDNA, a new safety dataset pairing literal translations with culturally localized prompts. They also propose a latent geometric framework to probe refusal representations in hidden states. Results show cross-lingual safety transfer is severely limited: harmful prompts retain less than 10% of the English refusal signal in most language-model pairs. Literal and localized prompts are semantically aligned (cosine 0.95–0.996) but drift across layers, meaning models encode concepts without routing them to safety mechanisms. The authors argue current multilingual safety alignment is superficial and challenge the notion of a universal, language-agnostic harm manifold in low-resource languages.

## Key points

- Cross-lingual safety transfer is severely limited; harmful prompts retain less than 10% of the English refusal signal across most language-model pairs.
- Introduces LoDNA, a safety dataset pairing literal translations with culturally localized prompts for Twi, Hausa, Amharic, and Swahili.
- Proposes a latent geometric framework that probes refusal representations in hidden states, going beyond generation-based evaluation.
- Literal and localized prompts are semantically aligned but drift across layers, indicating models do not route concepts to safety mechanisms.

## Full text

Safety alignment in large language models (LLMs) is largely developed in English, assuming these safeguards generalize across multilingual settings. However, this assumption remains underexplored and exposes a vulnerability in low-resource languages. We investigate cross-lingual safety transfer in four African languages, Twi, Hausa, Amharic, and Swahili, using LoDNA, a new safety dataset that pairs literal translations with culturally localized prompts. To move beyond generation-based evaluation, we propose a latent geometric framework that probes hidden-state refusal representations in LLMs. Our experimental results show that cross-lingual safety transfer is severely limited; harmful prompts retain less than 10% of the English refusal signal across most language-model pairs. Literal and localized prompts are semantically aligned (cosine 0.95-0.996) but drift across layers, suggesting models encode the concepts without routing them to safety mechanisms. These findings demonstrate that current multilingual safety alignment is superficial, providing strong evidence against the assumption of a universal, language-agnostic harm manifold within the specific low-resource languages studied. Warning: This paper contains example data that may be offensive or harmful.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_8ecaca672ff564d9de75e4c8