Who Holds the Key When AI Acts? A Hardware-Enforced AI Kill Switch and Execution-Time Gate for Machine Enforcement of the EU AI Act, US AI Law and Global AI Regulation

AI laws around the world now demand human oversight, purpose limitation, record-keeping, robustness, incident reporting and, increasingly, the ability to stop an AI system. Yet almost no technical layer enforces these rules at the one moment that matters: when an AI model’s or AI agent’s output becomes a real-world effect. Examples include a payment, a database write, a tool or API call, a message, a network change or a physical actuation. Today’s controls act before that moment (identity and access management, policy engines, guardrails, model evaluation) or after it (logs, audits, monitoring). That leaves a time-of-check to time-of-use gap that prompt injection, stolen credentials, model substitution and autonomous agents can exploit. This 61-page technical disclosure presents an architecture for machine enforcement of AI regulation built on one principle: computation is not authority to act. The authority to perform an operation is separated from the capability to perform it. Every consequential operation is non-completable by default. The capability it needs (a key, a release token, a credential, a network path or an electrical enable signal) is held in a protected, hardware-rooted enforcement domain such as a trusted execution environment, hardware security module, secure element or DPU, which the AI system cannot read or override. At execution time, the enforcement point jointly verifies five conditions: a scoped, revocable identity; a declared purpose bound in a signed policy token; a cryptographic fingerprint of the exact approved model and code; legal jurisdiction proven through certificate chains and attestation rather than IP addresses; single-use temporal validity. A signed, hash-chained receipt is committed before the capability is released or refused, so refusals are recorded as well as approvals. Anything uncertain fails closed to a defined safe state. Withholding the capability therefore works as an AI shutdown mechanism that stays effective even when the model or its host does not cooperate. The document contains more than thirty drawings, including regulation mapping drawings, a world map of AI regulatory approaches and a jurisdiction-by-obligation matrix. It maps the architecture article by article to: the EU AI Act as amended by the Digital Omnibus (Regulation (EU) 2026/1744), along with the GDPR, NIS2, the Cyber Resilience Act and DORA; US AI regulation, including state laws and a politically neutral analysis of proposed US federal AI legislation on frontier-model oversight, shutdown capability, suspension orders, pre-deployment approval, model-weight exfiltration, incident reporting and independent verification; the UK, China, India, South Korea, Japan, Singapore, Brazil and Canada; the OECD AI Principles, the Council of Europe AI Convention, the NIST AI RMF and ISO/IEC 42001. Sector maps cover finance, healthcare, employment, telecommunications, critical infrastructure, frontier models and AI agents, including MCP tool-call authorization. An industry-relevance section reviews publicly described AI agent security directions at Google, Microsoft, Meta, NVIDIA, AWS, Apple and frontier AI developers as independent convergence, without alleging copying or infringement. That section, and all regulatory and industry-relevance material, is AI-generated, and readers must cross-check it against the original PCT filings before drawing any conclusion. This disclosure is based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026 (WO 2026/150382), which claims priority from PCT/IB2026/053385, filed 7 April 2026 and publicly disclosed 4 June 2026 (WO 2026/115520). Related family applications are listed, and all numbers can be cross-checked on WIPO PATENTSCOPE. Licensed under CC BY-NC 4.0. The author is open to acquihire or strategic collaboration. Keywords AI kill switch; AI shutdown capability; machine enforcement; execution-time enforcement; runtime AI governance; EU AI Act; Digital Omnibus; US AI regulation; US federal AI legislation; frontier AI regulation; AI agent security; agentic AI governance; tool-call authorization; Model Context Protocol security; prompt injection; policy enforcement point; zero trust for AI agents; capability-based security; hardware root of trust; confidential computing; trusted execution environment; remote attestation; model fingerprinting; fail-closed design; tamper-evident audit log; AI incident reporting; model weight exfiltration; GDPR purpose limitation; data sovereignty; cross-border AI; human oversight; compliance as code; NIST AI RMF; ISO/IEC 42001; global AI regulation map; AI governance.