Charter — mission, scope, authority, exclusions

The charter is the first formal artifact of a new AI red team function. Without it, the team has no defined mission, no documented authority, no clear scope, and no defensible boundaries. With a vague one, the team will be misused (asked to evaluate a vendor’s marketing claims, asked to “test” a chatbot in production during business hours, asked to attribute incidents). A clear charter prevents all of this.

What a charter contains

Six sections. None optional.

1. Mission

One sentence. Example: “The AI Red Team validates the resilience of in-scope AI systems to adversarial inputs, with findings produced in a form suitable for executive decision-making, audit evidence, and regulatory examination.”

2. Scope (and what’s out of scope)

What does the team cover? Common scope inclusions:

• All AI systems above a documented risk threshold (typically EU AI Act high-risk equivalent, customer-facing, or systems handling regulated data).
• Pre-production, material-change, and periodic testing of in-scope systems.
• Evidence production for the AI Controls Catalog and applicable regulatory regimes.

Common scope exclusions (state these explicitly):

• Vendor evaluation as a procurement gate (that belongs to TPRM; AI Red Team may contribute technical evaluation).
• Operational SOC investigation of AI-related incidents (that’s IR; the AI Red Team contributes domain expertise during response, but does not lead).
• “Try to hallucinate this” requests — the team’s evaluation methodology defines what is tested; ad-hoc requests are triaged.

3. Authority

Who authorizes engagements? Typically the AI Governance lead and the system owner jointly. Major engagements require sponsoring-executive sign-off (CISO or CRO, depending on org). State that engagements are only conducted under written authorization, and that the team has no implicit authorization to act outside an authorized engagement — even in research mode.

4. RACI

Spell out:

• Responsible: AI Red Team lead.
• Accountable: CISO (or CRO if AI sits under risk).
• Consulted: AI Governance, Legal (for high-risk engagements), Privacy (when personal data is in scope), Communications (when public impact possible).
• Informed: Audit committee (quarterly), board (semi-annually).

5. Conflict of interest and ethics

The team does not test systems where members have a personal stake (built it, manages the team that built it). Inviolable hard limits — no production customer-data extraction, no actions that move real money, no targeting of specific named individuals — are reproduced from chapter 5.

6. Review cadence

Charter is reviewed annually and revised with sponsoring-executive approval.

Sample charter

A working draft is downloadable at /templates. Use the Engagement Planner (/planner) to generate a per-engagement Rules of Engagement that operationalizes the charter for a specific scope.