Asosiy kontentga oʻtish
Hisobot

Containment Is Not Authorization: Why a Sandbox Alone Does Not Establish a Runtime Authorization Boundary

Edward MeymanFerro (United States)
2026en
ABI

Annotatsiya

Version 1.0, September 2026. Author: Edward Meyman. Publisher: FERZ, Inc. Licensed under CC BY 4.0. This article follows The Medieval Way to Regulate AI (ferz.ai, June 25, 2026), which remains unchanged as a historical publication, and uses the Five Tests Standard specification v1.2.0. Containment restricts where a system can reach. Successful containment does not establish that the actions it permits are authorized. The article examines runtime authorization for consequential actions: an enforced pre-execution determination under the applicable policy and delegated authority, supported by an independently reconstructable authorization artifact, where execution requires a valid ALLOW and unresolved authorization remains blocked. It credits what containment and conventional access controls enforce, states what a perimeter alone does not establish under the three independent properties of the Authorization Boundary Integrity Model (Output, Input, and Replay Integrity), and applies the interruption distinction from the Impossibility result as an analogous evidentiary test: the fact that an action remained within a perimeter does not establish that it was permitted under the governing authority and policy. An architecture that demonstrates the stated properties satisfies the requirement regardless of its label. Two cases are distinguished. In the first, the perimeter is executed around. In the second, the perimeter holds and the system still exceeds its delegated authority through interfaces it was permitted to use; an illustrative messaging example, in which an approved channel carries both a permitted status update and an unauthorized confidential attachment, carries that case and answers the objection that better containment alone resolves it. The July 2026 evaluation incident involving OpenAI models and Hugging Face infrastructure is read as a coverage question on the three primary accounts published by the two companies: which consequential effects were subject to an enforced pre-execution verdict, and what evidence establishes that coverage. The incident demonstrates failures of isolation and access restriction; it does not, by itself, establish the absence of every authorization control. The article separates evaluation evidence from execution authorization, separates the origin of verdict-grounding inputs from the authority delegated to their source, and restates what authorization requires at a perimeter under the Five Tests Standard (Stop, Ownership, Replay, Escalation, Provenance), with an illustrative application of the five tests to the running example and no conformance claim. Related works in the FERZ corpus: The Authorization Boundary Integrity Model (v1.2); The Authorization Artifact Test (v1.2); On the Impossibility of Observability-Based Authorization (v1.4.0); The Five Tests Standard (5TS); A Taxonomy of AI Governance Approaches; The Closed-World Bargain; The Hook Is Not the Boundary; Technical Advisory TA-2026-01. The authoritative version of this article is maintained by FERZ, Inc. at ferz.ai; this Zenodo record is the timestamped record of publication.

Hali tarjima qilinmagan

Identifikatorlar

Iqtiboslar va manbalar

0 ta iqtibos0 ta foydalanilgan manba