How do AI authorization policies work in AI Control?
AI Control policies are structured, versioned rules: a set of conditions over fields such as agent, capability, environment, resource, target and risk, combined with an effect — allow, limit, require human approval, quarantine, revoke or block. When several policies match, the most restrictive effect wins. Policies are simulated, approved, signed, deployed to cloud and edge, and fully audited.
- Operators: eq, neq, in, gt, gte, lt, lte, contains.
- Most-restrictive-wins evaluation.
- Versioning with rollback.
- Signed bundles for edge enforcement.
Frequently asked questions
Can I test a policy before deploying it?
Yes. POST /api/v1/policies/simulate evaluates a draft against a sample request; deployment is a separate, audited step.
What happens when two policies conflict?
The most restrictive effect applies — for example BLOCK outranks REQUIRE_APPROVAL, which outranks LIMIT and ALLOW.