Principles

Each of these holds everywhere in awzy. The linked pages explain how.

How awzy decides

• Decisions are fast because the thinking happens first. Anything slow, such as classifying a document, scoring risk, or running an investigation, happens ahead of time and is stored with the person or resource. A check only reads what’s stored and evaluates policy, so it never waits on a model and takes milliseconds. See how awzy decides.
• One decision, whatever the path. A tool call through the MCP gateway and a request your own app sends to the decision API are decided by the same PDP and the same policies.
• An agent never gets more than its person. An agent acting for someone is checked as itself and as them, and the stricter verdict wins. See people, agents, and delegation.
• Watch before you enforce. A new install records what it would deny and challenge without applying it, and so do automated containments. You switch to enforce after reviewing them, with your security key. See monitor, then enforce.
• Ask before refusing. Where a flat deny would stop someone’s work, a challenge asks for a justification, a security key, or someone’s approval instead. See challenges and containment.
• Fail safe. A PDP that hasn’t loaded your policy refuses to decide rather than guess. When the PDP can’t be reached, the gateway denies, except for tools you’ve chosen in low-sensitivity tiers, and never for someone contained before the outage. See the MCP gateway.

Where AI fits

• Models can raise risk, never lower it. Risk scores and sensitivity come from rules first. An investigation can raise a score but never take it below the rules’ score, and a sensitivity label is never lowered. See risk and investigations.
• Agents propose; people decide. Agents can draft a policy or ask for a containment to be lifted. A person publishes and lifts, and only a person can pass a security-key challenge.
• Models run in your cloud. Awzy Bot and investigations use your own cloud’s model service, and are off until you turn them on. See optional features.

Your data and your install

• Your data stays in your account. awzy has no access to your install, even in an emergency, and only a fixed-schema heartbeat leaves it, which you can turn off. See the security model.
• Whatever changes enforcement is signed. Policy, the switch between monitor and enforce, and every image awzy ships. Your install applies only what verifies.
• There’s always a way back. Break-glass rollback works with keys your admins enroll at install, even if the IdP, the PDP, or the client portal is down. See break-glass.
• awzy follows its own rules. The client portal and its MCP endpoint are authorized by the PDP like any other app, and what agents do there is recorded like anywhere else.