Deep Learning Indaba 2026 · Lagos · Workshop

Building red-teaming capacity for AI safety evaluation in African contexts

This workshop introduces participants to red-teaming as a practical, hands-on methodology for probing models for harmful, unsafe, or misaligned behaviour, grounded in African languages and real-world use cases.

Format: Interactive workshop
Duration: Three hours
Capacity: Up to 40
Bring: A smartphone or a laptop
Reading list: Download PDF ↗

§ Schedule

Interactive workshop.

8:30 – 11:30am · Room: see Baobab app

Part I

Setting the scene

50 min

Part II

The method

40 min

Part III

Hands-on exercise

60 min

Part IV

Debrief & next steps

30 min

Understand why Africa-specific evaluation matters.

Can describe the methodology end-to-end.

Have run a red-teaming sequence yourself.

Part I Setting the scene Ground the room in why AI evaluation, and red-teaming within it, matters for African contexts.

8:30 – 8:40amWorkshop welcome and agenda
8:40 – 9:05amPanel discussion
9:05 – 9:20amAudience Q&A

Moderator: Joanna Wiaterek

Parts II & III The method, then the exercise Teach the red-teaming protocol end-to-end, then apply it hands-on in small mixed groups.

9:20 – 10:00amProtocol walkthrough
10:00 – 11:00amRed-teaming exercise

Session lead: Sumaya Nur Adan

With thanks to our facilitators for helping run the group exercise: Tamilore Oni, Chris Emezue, Maxwell Igweogu, Jabez Magomere and Sarah Luger.

Part IV Debrief & next steps Surface cross-group findings and share with participants further opportunities.

11:00 – 11:20amSharing of outputs
11:20 – 11:30amSumming up and next steps

§ Speakers - Hybrid Panel

Dr Laura Weidinger

Staff Research Scientist

Google DeepMind

JHO

Prof. José Hernández-Orallo

Director of Research & Research Professor

Leverhulme Centre for the Future of Intelligence, University of Cambridge

A/Prof Jonathan Shock

Associate Professor; Interim Director

University of Cape Town; UCT AI Initiative

§ The exercise

A guided red-teaming exercise.

In Part III you'll work in small mixed groups (technical and policy researchers together) to apply the methodology to a focused red-teaming exercise. Facilitators circulate to coach the work and push prompt construction.

Bring a laptop. Materials and facilitation guides are provided on the day.

Booklet Workshop overview Download PDF ↗ Worksheet Exercise step-by-step Download PDF ↗ Interface Red-teaming tool demo Watch demo ↗

§ Who it's for

Built for a mixed audience.

AI safety & NLP researchers

Working in low-resource or African-language settings.

Policy & social-science researchers

Working on AI governance, harm, or accountability.

Civil society & applied researchers

Evaluating AI tools being deployed in African contexts.

PhD & postdoctoral researchers

Looking for a methodological frame for safety work.

§ Responsible conduct

The exercise teaches a method, not how to generate harm.

The exercise is designed to teach red-teaming as an evaluation methodology. The facilitation cards, threat models, and annotated examples are prepared in advance by the organising team and are calibrated to support skill development without requiring participants to generate content that would be distressing to produce or encounter in a group setting.

01

Participant information

The workshop description published in advance states clearly what the exercise involves. Participation is voluntary, and an age requirement of 18 or above applies.
02

Written consent

Participants provide written consent at the start of the session covering participation in the exercise.
03

Facilitator oversight

All facilitators are briefed in advance on the protocol and on the procedures for redirecting group work if needed.
04

Data handling

No personally identifiable information is collected. Prompts are documented without attribution to individuals. Research outputs will not reproduce detailed adversarial prompts verbatim.

§ FAQ

Common questions.

Do I need to know how to red-team already?

No prior experience is required. The workshop is designed to introduce the methodology from scratch, with facilitators guiding you throughout the hands-on exercise.
Will I leave able to red-team independently?

You'll leave with a solid working understanding of the method and have completed a full red-teaming sequence. Independent practice will develop further with time, and we'll share resources to support that.
I work in policy and don't code — is this for me?

Yes. The workshop is explicitly designed for mixed technical and non-technical audiences. Groups are intentionally mixed, and the methodology does not require coding skills.
How do I sign up?

Registration details will be announced alongside the Deep Learning Indaba 2026 programme. Check back here or follow CASA and UCT AI Initiative for updates.

§ Funders

This workshop is funded by the Centre for AI Security and Access (CASA), including facilitation and delivery costs. We gratefully acknowledge further key support from:

Cohere Labs' Catalyst Grant — API credits enabling participants to work with the Aya and Command model families.

The UCT AI Initiative's contribution to this work is supported by the Artificial Intelligence for Development (AI4D) Africa programme, with financial support from Canada's International Development Research Centre (IDRC) and the UK's Foreign, Commonwealth & Development Office (FCDO).

The views expressed herein do not necessarily represent those of IDRC or its Board of Governors, or those of FCDO.