Chapter 7 — Red Teaming with AI & Humans in the Loop
How to red team when AI assists analysis and humans retain decision authority — grounding, dissent, quality control, and hybrid session design.
Concepts & tools in this chapter
Section titled “Concepts & tools in this chapter”- AI-Assisted Red Teaming — Use AI as loyal opposition — challenge plans and surface risks — while the human owns stakes, judgment, and commitment.
- Human-in-the-Loop Decision Authority — The decider remains accountable. AI informs and challenges; it does not approve, commit, or replace stakeholder judgment.
- Grounding & Evidence Discipline — Read actual source material before analyzing. Separate grounded claims from model invention, guesswork, and phantom citations.
- AI Check — Meta-review: mirror, grounding, dissent, sycophancy on AI-assisted analysis.
- AI Anti-Patterns — Sycophancy, false balance, phantom citations, shallow steel-man, checklist theater in AI output.
- Hybrid Session Design — Allocate roles across AI and humans across diverge → analyze → debate → converge with explicit handoffs.
- Simulated Stakeholder Limits — When invented stakeholder voices help vs mislead — label simulations; prefer real dissent and
culturewith actual actors.
RedTeam commands
Section titled “RedTeam commands”/redteam ai-check/redteam challenge/redteam review/redteam steelman/redteam calibrate/redteam record/redteam sequence
RedTeam note
Section titled “RedTeam note”Handbook v10-native. UFMCS v9 predates modern AI-assisted workflows; this chapter defines how the RedTeam skill should behave with humans in the loop.