Abstract

Trust between the United States and China on AI is limited, and that limit holds back cooperation on AI safety. Neither side is willing to disclose model weights, training data or internal evaluations to the other. Incident hotlines and shared testing communicate or measure after the fact. Most technical controls assess the model rather than the moment an AI-initiated operation actually takes effect. Yet real harm occurs at that moment, when a payment settles, a network configuration changes, an actuator moves or a software release ships. This gap is the question of who controls the final action.

Based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026, and Indian Provisional Filing 202631136555, dated 15 November 2026.

This paper presents a technical framing of that gap using an architecture for consequence-bound execution authority. Permission to compute, propose or prepare an operation is separated from permission to cause its real-world effect. A proposed operation is held in a pending state. The operation that would actually execute is reconstructed. An independent policy decision point checks current authority, context, dependencies and revocation status. A scoped, non-bearer, single-use authorization is issued, and a protected enforcement point at the commit boundary re-derives the operation, enforces that authorization and records evidence for later reconciliation.

The U.S.–China relevance is that when trust is limited, cooperation can rest on verification of outcomes instead of trust in the other side’s models. Each jurisdiction keeps its own policies, keys, infrastructure and final authority over consequences. The shared understanding concerns one verifiable property (no consequential effect without current, checkable authority) rather than access to model internals. The approach is intended to complement incident hotlines, red-team evaluation and standards work, not replace them.

The scope is high-criticality systems and high-consequence operations, where bounded latency is acceptable but catastrophic or irreversible harm is not. It does not address microsecond-scale safety loops. No performance, latency or adoption claim is made. It is not a treaty mechanism, a standard, a regulatory compliance statement or a claim that either government would adopt it. It credits existing identity, authorization, attestation and confidential-computing controls and is designed to layer with them.

Industry relevance (proposed integration points, not implementations). In each case the enforcement point is added at the point of effect and the vendor’s existing controls stay authoritative:

  • AI agent platforms. In the OpenAI Agents SDK, the tool dispatcher re-derives each consequential function call, checks an authority record held outside the prompt, and rejects calls whose target or scope differs, so prohibitions are not lost in handoffs. For Anthropic Claude tool use and MCP, the MCP host looks up the authority contract for each tool call and the server verifies a scoped capability against the actual operation, rejecting widened scope. In Microsoft Azure AI Foundry agents, the managed identity is used only by the enforcement point after verification, so an agent cannot reach the target with its own credential. In Google Cloud agent frameworks (ADK, A2A), the composed result of a multi-agent run is validated at the shared commit step and the authority contract is preserved across agent hand-offs. For AWS agent services, end-target checks cover paths a gateway cannot see, such as provider-hosted or direct-credential routes, and rejection evidence shows which routes are covered.
  • Compute and confidential computing. For NVIDIA GPUs and BlueField DPUs and for AMD Instinct and Pensando, the output of a failed-over workload is cut over to a single writer, and the installed offload state is read back and compared with the authorized flow. For Intel TDX, attestation evidence is joined with patch state and policy before an effect is admitted, and old trust-domain paths are closed on reassignment. For Arm CCA, a domain transition controller withholds successor authority until DMA and interrupt paths are closed. For Qualcomm, Apple and Samsung devices, stored intents are revalidated at wake-up and output is released only to the intended audience across display and casting routes.
  • Telecommunications. For Ericsson and Nokia network automation (rApps, SMO), the check sits where an application’s output becomes an applied configuration: multi-cell and slice effects are reconstructed, validated, applied as a limited change and expanded only on attributable impact evidence. This guards against a locally successful change that harms neighboring cells. Real-time radio scheduling is untouched.
  • Enterprise AI and data platforms. For SAP, a validator at the posting boundary recomputes tax, currency and allocation effects from live state and compares them with the approval before the posting commits. For Oracle databases, the full set of effects reachable from a stored program (triggers, definer-rights code, autonomous transactions) is enumerated and each material effect needs authority. For Salesforce Agentforce, a “completed” message to a customer is released only after authoritative fulfillment evidence exists. For ServiceNow, the live dependency topology is revalidated at the remediation commit. For IBM watsonx, sink-rejection evidence per route turns governance dashboards into proof of coverage. For Databricks Unity Catalog and Snowflake Cortex, tests check that data-access obligations survive every query route and that the released result matches what was authorized.
  • Finance. For Visa, Mastercard, Swift, Adyen and Worldline, an exposure ledger and reservation sit between authorization and capture, beneficiary-net checks run on cross-border flows, and retries are quarantined after lost acknowledgments, so an agent-initiated payment cannot exceed its authorized exposure. For the ECB’s digital euro research on offline payments, the proposal is a research extension for delegated child allocations and reconciliation after compromise, with existing secure-element controls remaining authoritative.
  • Industrial systems. For Siemens and Schneider Electric edge platforms, deployments and remote commands are gated at the edge runtime and the actuation boundary, local safety interlocks are kept, digital-twin results are used as evidence and not as authority, and an unavailable validator maps to the system’s safe state.

This industry-relevance material is AI-generated and consists of the author’s engineering hypotheses. It is not an endorsement, validation or infringement statement, and it does not show that any company implements or intends to implement the proposed combinations. Product descriptions must be re-verified against current vendor documentation.

Related publication: “Europe’s Next AI Security Challenge: Who Controls the Final Action?” (Zenodo, DOI 10.5281/zenodo.24274259).

Where to read the full PCT filing: search “Sangam Das” on WIPO PATENTSCOPE; search PCT/IB2026/055615 (pctib202655615); or read the PCT file directly in TD Commons at https://www.tdcommons.org/dpubs_series/12542/.

Disclaimer: This abstract and the drafting were prepared with AI assistance. Readers must cross-check them against the original disclosures, the PCT filings and other public sources before drawing any conclusion.

Keywords 

AI Safety, Trustworthy AI, U.S.–China AI Cooperation, Limited-Trust AI Collaboration, AI Governance, Agentic AI Security, Autonomous AI Agents, AI Agent Security, Autonomous Systems Security, Runtime Authorization, Policy Enforcement Point (PEP), Policy Decision Point (PDP), Zero Trust Architecture, Reference Monitor, Complete Mediation, Non-Bypassable Execution Control, Consequence-Bound Authorization, Protected Execution-Finality, Critical Infrastructure Protection, Critical Infrastructure Security, National Security, Defense Cybersecurity, Sovereign AI, Digital Sovereignty, AI Incident Response, AI Risk Management, High-Risk AI Systems, Human Oversight of AI, Confidential Computing, Trusted Execution Environment (TEE), Hardware Root of Trust, Remote Attestation, Cryptographic Attestation, Capability-Based Security, Non-Bearer Token, Sender-Constrained Authorization, Least Privilege, Confused Deputy Prevention, Time-of-Check/Time-of-Use (TOCTOU) Prevention, Fail-Closed Enforcement, Revocation Checking, Secure Command Authorization, Independent Effect Verification, Tamper-Evident Audit Evidence, Evidence-Carrying Receipts, Outcome Reconciliation, Secure Multi-Agent Systems, Agent-to-Agent (A2A) Security, Model Context Protocol (MCP) Security, Tool-Call Authorization, Function-Calling Security, Cross-Platform AI Agent Security, Cross-Domain Security, Cyber-Physical Systems Security, Operational Technology (OT) Security, Industrial Control Systems Security, Mission-Critical Systems, UAV and Autonomous Vehicle Security, Robotics Safety, Space and Satellite Systems Security, 6G Network Security, Network Slicing Security, AI-Native Networks, Open RAN Security, Digital Payments Security, Agentic Commerce Security, Central Bank Digital Currency (CBDC), Digital Euro, EU AI Act, NIS2, Cyber Resilience Act (CRA), DORA, NIST AI Risk Management Framework, ISO/IEC 42001, IETF Internet-Draft, Patent-Pending Technical Disclosure, Technical Disclosure Commons, Consequence-Bound Execution Authority, Finality Sink, Operation Reconstruction, Actual Pending Operation, Candidate Act, Non-Effective State, Source-Authority Envelope, Protected Enforcement Domain, Independent Validator, Non-Bearer Execution Capability, Effectuation-Boundary Enforcement, Control-to-Effect Route Coverage

License: CC BY-NC 4.0.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Share

COinS