Request for Proposal Independent Red Teaming and AI Safety Evaluation

Procurement Summary

Item Details

RFP Title Independent Red Teaming and AI Safety Evaluation

Procurement Method Open competitive tender

Assignment Type Independent organisation or specialist consultancy

Primary Market Kenya

Contract Period Approximately 10 weeks

Reporting To Head of Technology and Product

Clarification Deadline 10 November 2026

Proposal Deadline 18 November 2026

Submission Email [email protected]

Independence is a mandatory condition. The selected evaluator must not be involved in building, operating, or commercially benefiting from the systems evaluated under this assignment.

Who We Are

Girl Effect is an international non-profit that builds media and technology that adolescent girls want, trust and need. We create digital-first brands and products that reach girls at scale across Africa and Asia, supporting them to make informed choices about their health, education and futures — with a particular focus on sexual and reproductive health (SRH).

Girl Effect has worked in responsible AI for several years and has published its own Ethical AI Guidelines and a Generative AI Whitepaper, co-authored sector guidance on the safe use of chatbots for social and behaviour change, and operates named accountability for AI safety — including a Failure Mode Register with safeguarding sign-off, and defined escalation protocols with service-level commitments for human follow-up. We believe that AI serving young people in the Global South must be held to the highest evidential standard, and that the organisations building such systems should not be the only ones marking their own homework. This engagement exists to put that principle into practice.

Background

WAZZII is Girl Effect’s WhatsApp-based chatbot in Kenya. It provides adolescent girls and young women with confidential and non-judgemental information about their bodies, relationships and health. It also connects users to relevant products, referrals and health services, supported by a human help desk responsible for moderation and escalation.

WAZZII’s next phase will expand the use of generative AI in the girl-facing experience. This includes:

Retrieval-grounded responses using clinically approved content;

AI-assisted tools for human help-desk operators;

Safeguarding and clinical escalation pathways;

Internal evaluation of model outputs; and

Defined policies covering content and actions that the system must never produce.

Girl Effect seeks an independent partner to test whether these safeguards perform as intended and identify risks that may not have been detected through internal evaluation.

The selected partner must be separate from the organisations developing or operating the systems being tested.

Objective

The objective is to conduct a focused and independent red-team evaluation of WAZZII’s girl-facing generative AI capability.

The assignment will:

Identify safety, safeguarding, clinical and technical vulnerabilities;

Test whether the system recognises and safely escalates relevant risks;

Assess whether existing evaluations and controls adequately reflect real-world use;

Produce reproducible test cases that Girl Effect can rerun;

Rank findings by severity, likelihood and potential impact; and

Provide practical and prioritised recommendations for remediation and future evaluation.

The core assignment is a single time-bound evaluation. Remediation re-testing, review of agentic designs and periodic testing will be priced separately as optional extensions.

What You will do

The core engagement runs over an indicative ten weeks. All work follows Girl Effect’s Ethical AI Guidelines, data policy and safeguarding policy: testing is conducted against staging environments using synthetic personas, no real user data is accessed or collected, data collection is minimised throughout, and no safeguarding case content is handled at any point. Girl Effect will provide staging access, including programmatic (API) access to the conversational endpoint, supporting automated test harnesses; testing may be conducted fully remotely, with no in-country presence required. Key activities include:

Phase 1 — Orientation and test design (indicatively weeks 1–3)

Focused desk review of Girl Effect’s evaluation framework, red-line policies and escalation protocol — sufficient to design a well-targeted test plan, not a standalone audit

Co-scoping of the scenario taxonomy and severity rubric with Girl Effect’s AI Ethics Officer and safeguarding leads, and agreement of the responsible-disclosure protocol for critical findings

Phase 2 — Red-teaming sprint (indicatively weeks 3–8)

A structured, tool-supported testing exercise against the girl-facing GenAI experience in staging, combining adversarial and realistic non-adversarial scenarios and producing logged, machine-readable, re-runnable test cases, prioritising: clinically incorrect or harmful health advice; failure to recognise and escalate safeguarding disclosures; jailbreaks, prompt injection and policy circumvention; and age-inappropriate content

Testing across the languages and code-switching patterns of real users (including Kiswahili and Sheng); local-language competence may be provided by the partner or in pairing with Girl Effect's own team

Stress-testing of the human-escalation pathway, primarily with realistic non-adversarial scenarios (genuine distress, disclosure and clinical-concern personas where the system should hand over without being attacked): does it hand over to the help desk when it should

Phase 3 — Findings and briefing (indicatively weeks 8–10)

A severity-ranked findings report with reproducible test cases, handed over so Girl Effect’s team can re-run them continuously after the engagement ends

A findings briefing for Girl Effect leadership, with prioritised recommendations — including what a fuller evaluation should cover next and at what cadence

Optional priced extensions (quoted separately, commissioned at Girl Effect’s discretion)

Re-test of remediated critical findings

Architecture-level adversarial review of agentic system designs (threat-modelling autonomy, tool use and failure modes before deployment)

Periodic re-testing and pre-launch adversarial gates for major releases

Deliverables

Deliverable

Description

Indicative timing

Inception note & test plan

Confirmed methodology, scenario taxonomy and severity rubric, staging/access arrangements, ethics and data-handling approach, responsible-disclosure protocol

Week 3

Red-team findings report

Severity-ranked, reproducible findings from the adversarial sprint, including escalation-pathway results; test cases handed over for continuous internal re-use

Week 8

Leadership briefing & recommendations

Findings briefing with prioritised recommendations and proposed scope/cadence for future independent evaluation

Week 10

Duration and Working Arrangements

The core engagement is expected to run for approximately 10 weeks.

Testing may be completed remotely. No in-country presence is required unless separately agreed.

Girl Effect will provide:

Approved access to the staging environment;

Programmatic access to the conversational endpoint where available;

Relevant documentation and evaluation materials;

Named technical, safeguarding and programme focal points; and

Consolidated feedback on deliverables.

The partner will:

Participate in a weekly progress meeting;

Maintain a shared delivery tracker and risk register;

Use secure communication and evidence-sharing channels;

Escalate critical issues immediately;

Protect access credentials and evaluation materials; and

Manage its personnel and delivery approach independently.

Independence and Conflicts of Interest

Independence is a mandatory condition of award.

The bidder must:

Have no current contract to build or operate the system being evaluated;

Have no commercial interest in the evaluation achievi