AI Feature Toggles

The top of Settings → AI shows four switches: AI Inference, AI Agents, AI Autonomy, and AI Training. They are organization flags, not per-user preferences. Customers see the current state but cannot flip them — the switch is disabled with Managed by Praetorian. Only a Praetorian operator (or the admin@praetorian.com operator account) with the Admin (manage_settings) entitlement can change them. The API also rejects non-Praetorian writes.

These flags do not configure models, spend, or hunt profiles. They turn whole classes of AI behavior on or off for the tenant.

The four flags

AI Inference — human in the loop. When on, the AI assistant can read tenant data and answer questions. You still decide what to do with the answer. Enabling it confirms that the assistant may process queries using this account's data.

AI Agents — human on the loop. When on, the assistant may act (triage, validate seeds, launch approved work) while you supervise. This is a positive flag: it is off until explicitly enabled. It requires AI Inference to already be on. If Inference is off, Agents stays locked with Enable AI Inference first.

AI Autonomy — human off the loop. When on, admin and analyst users on this account can see hunt sessions, and Hannibal can run fully autonomous penetration tests. Read-only users still cannot see hunts; visibility is enforced on the server (flag and analyst/admin role). Autonomy does not depend on Inference or Agents — it is a separate org flag.

AI Training — when on, account data may be used to improve model performance, subject to Praetorian's data-handling policy. It requires AI Agents (and therefore Inference). Disabling Agents also disables Training.

Dependency chain

Inference → Agents → Training is a chain. Autonomy sits beside it.

Disabling a prerequisite cascade-disables dependents so they do not silently come back when the prerequisite is turned on again. Example: turning off Inference also turns off Agents and Training in the same save.

Enabling a flag that has a confirmation dialog (all four do) asks you to confirm before the write. Disabling does not.

What this is not

These switches are not the same as AI Model Providers (which vendors Guard may call), Monthly AI Spend (the cap), or Automatic Penetration Tests (the Hannibal schedule). Those cards still appear on the tab even when a flag is off; the schedule card becomes read-only when Autonomy is off.