In this challenge, participants will design and submit adversarial prompts that expose trust and safety weaknesses in large language models (LLMs). Submissions are a Markdown (.md) file containing one or more attack entries, where each attack includes:

• an adversarial prompt
• the model’s response
• structured metadata describing the attack

Participants must follow the structure of the Sample Submission and are provided with structured taxonomy files to ensure consistent labeling within your submission file.

Note:

• Each attack must include both prompt versions (original + English)
• The model response must support the claimed risk
• Metadata must align with the observed behaviour
• Submissions may contain multiple attacks, but quality is more important than quantity

🌍 Target Models & Supported Languages

This challenge focuses on evaluating trust and safety risks across a diverse set of African languages and language models. Participants are expected to design adversarial prompts in any of the supported languages below and evaluate model responses accordingly.

🧠 Available Models by Language

Swahili https://huggingface.co/sartifyllc/Pawa-Gemma-Swahili-2B

Hausa https://huggingface.co/NCAIR1/N-ATLaS

Yoruba https://huggingface.co/NCAIR1/N-ATLaS

Amharic https://huggingface.co/EthioNLP/Amharic_LLAMA_our_data

Igbo https://huggingface.co/NCAIR1/N-ATLaS

Oromo https://huggingface.co/EthioNLP/Amharic_LLAMA_our_data

Fulfulde https://huggingface.co/bonadossou/afrolm_active_learning

Pulaar https://huggingface.co/bonadossou/afrolm_active_learning Serere https://huggingface.co/bonadossou/afrolm_active_learning

Somali https://huggingface.co/EthioNLP/Amharic_LLAMA_our_data

Zulu https://huggingface.co/lelapa/InkubaLM-0.4B

Shona https://huggingface.co/bonadossou/afrolm_active_learning

Lingala https://huggingface.co/bonadossou/afrolm_active_learning

Afrikaans https://huggingface.co/lelapa/InkubaLM-0.4B

Wolof https://huggingface.co/bonadossou/afrolm_active_learning

Akan https://huggingface.co/Ghana-NLP/abena-base-akuapem-twi-cased

Tigrinya https://huggingface.co/EthioNLP/Amharic_LLAMA_our_data

Malagasy https://huggingface.co/bonadossou/afrolm_active_learning

Notes

• Some models support multiple languages (e.g. N-ATLaS, AfroLM, EthioNLP models)
• Participants may submit attacks across multiple languages and models
• You are encouraged to explore language-specific vulnerabilities
• Consider how safety behavior may differ across: translation tasks cultural context low-resource vs high-resource languages

This challenge aims to surface real-world trust & safety gaps in multilingual AI systems, particularly in underrepresented languages.