Designed so the strictest review starts bored.

In its read-only posture AI Pulse deploys as a managed application inside your own Azure tenant. It authenticates with a read-only managed identity; there are no write permissions to grant, scope down, or accidentally leave broad. Asset data, telemetry and generated evidence stay in your tenant — nothing is shipped to us to “improve the product”.

With enforcement deployed it runs where your agents run — any cloud, self-hosted, sovereign, or fully air-gapped. Its decision ledger is tamper-evident and stays in-environment. Enforcement power is granted per agent, by you, and can be read back as policy at any time.

An agent is only as governable as its identity. Agents authenticate either with a per-agent secret or with a workload identity token — verified signature, exact issuer and audience, expiry, and a binding claim that must resolve to that exact agent in the registry. A token that fails never falls back to a secret.

Moving from shared secrets to per-workload identity is a migration, so it runs in dual mode and stays off until you deliberately enable it. Two rules hold throughout: the shared platform identity is forbidden as an authentication subject, and a decision record stores only which mechanism was used — never a token, a claim or a secret.

Rotation stages new values under separate names on the services already entitled to hold them, changing nothing live until every holder has accepted its staged value. An interrupted run resumes from its checkpoint and replays the same values rather than generating new ones. Values are never printed; progress is reported as counts, and a failure names the holder and the secret it was missing.