The manifesto

Test what the system does—not what the policy says.

AI assurance must move from documents and intentions to observable behavior under regulatory pressure.

Why RegTeaming exists

AI systems increasingly make, recommend, explain, prioritize, collect, disclose and transact. In regulated settings, the important question is not only whether these systems are safe. It is whether their behavior remains lawful across users, products, jurisdictions and changing expectations.

A system can produce a helpful answer and still cross into regulated advice. It can complete a workflow while omitting a required disclosure. It can treat each step as acceptable while producing an impermissible overall outcome. It can pass a review and later drift as models, prompts, tools or regulations change.

Red-teaming asks whether a system can be made unsafe. RegTeaming asks whether it can be made unlawful—and whether the organization can prove it tested for that possibility.

Principles

1. Test behavior, not declarations.

Policies, controls and model cards express intent. RegTeaming tests whether the deployed system follows that intent under realistic pressure.

2. Link every test to authority.

Every regulatory evaluation should identify the governing source, obligation, scope, condition and expected behavior. A failure without provenance is difficult to interpret; a legal claim without a test is difficult to operationalize.

3. Test trajectories, not isolated answers.

Regulatory failures often emerge across multiple turns, tool calls, handoffs and workflow decisions. The unit of evaluation is the complete consequential trajectory.

4. Attack context.

Jurisdiction, product, customer vulnerability, language, channel and timing can change the applicable obligation. RegTeams systematically vary these conditions.

5. Test control effectiveness.

The existence of a disclosure, escalation rule or human-review policy does not prove it works. RegTeaming evaluates whether controls trigger correctly and change the final outcome.

6. Preserve evidence.

Each evaluation should retain the inputs, context, system version, trace, outcome, governing obligation and remediation status needed for review.

7. Update continuously.

Regulatory assurance is not a one-time certification. Tests must change when the law, regulatory expectation, model, prompt, product or workflow changes.

8. Report scope honestly.

Passing a defined test suite is evidence about specified behavior under specified conditions. It is not a universal declaration that a system is compliant.

An open practice

RegTeaming is intended to be used and extended by engineering, legal, risk, audit, assurance and research teams. Shared terminology and test structures make cross-functional regulatory assurance possible.

Use the method. Adapt it to your sector. Publish attack patterns. Improve the specification. Cite the obligations. Share what works.