AI Red-Teaming PlatformJoin an event— Closing the Gap Between Access and Security
Probe AI systems for safety failures, under structure.
Red-teaming — structured adversarial testing to surface an AI system's flaws before deployment — has so far been concentrated among a handful of labs and languages. CASA runs it as a live, facilitated exercise for researchers, policy practitioners, and civil-society technologists who need that capacity themselves, and hands them the findings to act on.
Method
Structured testing, not open-ended chat
Coverage
Every harm category tracked live, gaps visible mid-session
Data
No PII collected; retention set per event
Access
Public link for open events, invite code for closed ones
§ 01 — How It Works
Every session follows the same structure.
Part I
Consent & briefing
Participants confirm participation, opt out of categories they'd rather not see, and receive the scenario brief.
Part II
Probe the scenario
Each participant chats directly with a configured target system, in the language of their choice.
Part III
Flag findings
A flagged response opens a structured form: category, severity, scope, and a proposed fix.
Part IV
Review & report
Facilitators watch coverage live; CASA compiles a disclosure report after the session ends.
§ 02 — One Platform
Built to be reconfigured, not rebuilt.
Every event is its own configuration — a different model backend, scenario set, harm taxonomy, and language list — run on the same platform CASA maintains between deployments.
Live — Example Event
Deep Learning Indaba 2026 · Lagos
Invite-only · ~40 concurrent participants · 3 scenarios configured
§ 03 — Roles
Four roles, one shared record of findings.
Admin
CASA team. Configures events, watches coverage live, exports the disclosure report.
Red-teamer
Workshop participant. Probes a scenario, flags what breaks, moves on.
Annotator
Coming to a future release
Reviews flagged conversations and rates harm, blind to the participant's own tag.
Arbitrator
Coming to a future release
Resolves disagreements between annotators on contested findings.
§ 04 — Responsible By Design
Built to find failures, not manufacture them.
Scenarios and taxonomies are prepared in advance by the organising team. Participation is voluntary and reversible at every step.
01
Informed consent
Every participant sees what the exercise involves before starting, and confirms it in writing.
02
Content opt-outs
Harm categories can be opted out of before starting, and updated at any point via Take a break.
03
No PII collected
Conversations are recorded without attribution to individuals. Retention windows are set per event.
04
Facilitator oversight
Every session runs with a briefed facilitator able to redirect or pause the exercise.