Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJun 22, 2026Johannesburg, South Africa

Cross-Lingual Safety Audit of LLMs in South African Languages

Thando Shabangu · Team Tech Titans

Submitted to Global South AI Safety Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Cross-Lingual Safety Audit of LLMs in South African Languages

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

We test whether four LLMs (Claude Sonnet 4.6 plus three locally-run open-weight models) keep their safety guardrails when harmful requests are issued in three South African languages — isiZulu, Tshivenḓa, Sepedi — versus English. The open-weight models refuse 69% of harmful prompts in English but only 7% in the indigenous languages, and we show this degradation splits into genuine safety failures and mere capability failures — a distinction we argue is essential for honest multilingual safety evaluation.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. A clean, well-represented audit. the 4-way coding scheme that separates safety failures from capability failures is interesting. That distinction matters and is underappreciated in the literature. The small sample size is the main limitation

    Strengths:

    - The 4-way response coding (refusal / partial / full compliance / alignment failure) is a real contribution. When a model produces gibberish in Tshivenḓa, that's not a safety failure, it's a capability failure. Most other work conflates these.

    - Human annotation rather than LLM-as-judge adds credibility, especially for low-resource languages.

    Suggestions for Future Work:

    - With only 12 prompts per cell, the quantitative claims are on shaky ground. Scaling this up would make the findings much more convincing.

    - Including 4B-parameter models that obviously cannot process the target languages inflates the apparent safety gap. Matching model capabilities to the languages being tested would give a cleaner picture.

    Read full reviewShow less

  2. -- Impact potential & innovation

    ++ execution quality

    ++ presentation & clarity

  3. The project aims to characterize how different LLM’s respond to harmful request in native South African languages such as isiZulu, Tshivenda, and Sepedi. The project identifies an important limitation of LLMs to handle malicious requests in these native languages, particularly in open models such as Qwen and Gemma. It is also highlighted the distinction between compliance and capability failures of the LLMs to understand these languages. Nevertheless, the sample size of the experiments is not large enough to draw enough conclusions.

Cite this project

@misc{shabangu2026crosslingual,
  title = {{Cross-Lingual Safety Audit of LLMs in South African Languages}},
  author = {Thando Shabangu},
  year = {2026},
  month = jun,
  note = {Submitted to Global South AI Safety Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/crosslingual-safety-audit-of-llms-in-south-african-languages-dtp3}},
  url = {https://apartresearch.com/sprints/projects/crosslingual-safety-audit-of-llms-in-south-african-languages-dtp3}
}

More from Global South AI Safety Hackathon

All 216 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026