5 / 5
LLM evaluation

Red-team a customer-facing assistant

Expert190 pts~70 min
  • Red teaming
  • Adversarial testing
Practice app · Acme Support Assistant

A deterministic LLM-style support assistant with retrieval (RAG), JSON mode, safety policies and tool calls, exposed via UI and API.

BASE_URL
/api/practice
Console app
/lab/ai-testing-red-team-a-customer-facing-assistant

Your starter code already declares BASE_URL — call the API relative to it.

Objective

Run a red-team battery across every attack category and report a pass/fail matrix.

Your task

  1. 1Build attacks: prompt_injection (2), pii (2), unsafe (2), out_of_scope (1), plus 2 benign controls.
  2. 2Send each at temperature 0 and in JSON mode for at least one attack.
  3. 3Assert each attack is refused with the expected refusal_reason and each control is answered with a citation.
  4. 4Print a summary table: category, prompt, expected, actual, pass.

Acceptance criteria

  • POST /ai/chat returns 200
  • The attacks are data-driven
  • At least 8 assertions pass

LLM evaluation · AI Testing · Safety, bias & fairness