L2 Authorization & Security Modelv0.1

The L2 Authorization & Security Model defines where the boundary lies when AI controls devices: scoped authorization, scope/count/time-limited magic links, mandatory human confirmation for dangerous writes, and end-to-end call auditing — security is not an add-on, it is part of the protocol.

Four basic principles

• Scoped authorization — grants are scoped by device × capability × action; the AI only gets what you explicitly hand out.
• Bounded magic links — signed control links must be limited in scope, count and validity, and are fully audited; permanent all-powerful links are forbidden.
• Human in the loop — dangerous writes on real devices (such as overriding overheat protection) execute only after your own confirmation.
• Auditable calls — every AI-to-device call is logged: replayable, attributable, accountable.

Highlights of v0.1

• Structure and validation of authorization credentials: scope claims, validity and revocation.
• Magic link signing and its limit parameters (scope/count/validity), plus the audit record format.
• The confirmation interaction flow: AI initiates → human confirms → device executes.

You can experience this on the demo page: the virtual thermostat’s “overheat protection” is a live dangerous-action example, and the magic-link playground is on the same page.