Abstract

AI agents now do real work. They update business records, send files, move money, use connectors and run inside cloud virtual machines. Knowing who the agent is does not tell us whether this exact action, with these exact values, should be allowed right now. An agent can be fully authenticated and still send an edited payment, forward the wrong file or replay an old approval. That is an important AI safety gap.

This disclosure is based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026. That application is pending and has not been granted. Its pattern is that an AI action waits as a pending operation. An isolated policy decision point rebuilds the real action and tests it against policy and the original delegated authority. A narrow, single-use, sender-constrained capability is issued. A policy enforcement point at the real effect verifies that capability before the action can commit. If anything is missing, changed or expired, the action is refused (fail-closed).

The disclosure compares this pattern with public 2026 descriptions from:

  • Microsoft: Agent 365, Entra Agent ID, Copilot Studio, Foundry, Azure API Management, Defender, Purview, Windows Agent Workspace, Dynamics 365
  • Meta: Muse, Hatch, Sentinel, PrivacyLib, AgentLib
  • Google: Agent Compute Engine, Agent Substrate, Agent Identity, Agent Gateway, Agent Registry, Model Armor

It also maps possible alignment with other large vendors and public product areas:

  • Amazon (AWS): agent identity and gateways, IAM, Nitro System
  • Anthropic: Claude agents and Model Context Protocol connectors
  • OpenAI: function calling and agent products
  • Apple: Secure Enclave and Private Cloud Compute
  • Salesforce: Agentforce
  • ServiceNow: AI agents and workflow automation
  • IBM: watsonx.governance
  • NVIDIA: NeMo Guardrails and confidential computing
  • Cisco: AI Defense

These platforms already have strong identity and security controls, so the disclosure asks a narrow question: does one extra exact-operation check add value at selected boundaries? It includes flow diagrams, fifteen possible new technical effects and worked examples, such as blocking an edited payment, stopping a read-only permission from being used to send email, and rejecting a replayed refund after a virtual machine snapshot restore. Checks apply only to consequential actions, so everyday AI work stays fast. All speed figures are engineering targets, not measured results.

The industry-relevance material is AI-generated. Readers must cross-check it against the original PCT filings before drawing any conclusion. The alignment shown is possible alignment only. It is not an endorsement, validation or infringement statement, and no company named is claimed to have implemented, copied or infringed anything. The author is open to an acquihire or strategic collaboration.

Full PCT filing: search “Sangam Das” on PATENTSCOPE, or search PCT/IB2026/055615 (pctib202655615). The PCT file can also be read directly in TD Commons: https://www.tdcommons.org/dpubs_series/12542/

Keywords: AI safety; AI agents; Model Context Protocol; policy enforcement point; sender-constrained capability; confidential computing; API gateway; credential isolation; fail-closed; TOCTOU

Licence: 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