Abstract

Companies are connecting AI assistants such as Claude and ChatGPT to email, files, code, customer records, finance and supplier systems. Each connection is approved on its own, but one AI that can see them all can connect the dots. Customer complaints, engineering tickets, code changes, supplier orders, new hires and budgets, read together, can reveal a company’s unreleased products and future plans, even though no single file says so. If that AI is hacked, tricked by a poisoned document, or simply over-connected, years of carefully separated security can collapse into one view that can be leaked or acted on. Today’s protections, such as permissions, contracts, filters and audit logs, decide who may open which file. They do not stop an AI that can read A and B from combining them and sending out the result. That is why banks, defence firms, hospitals and governments still hesitate to give AI deep access, and why companies hesitate to share data through AI.

This disclosure replaces promises with cryptographic locks. “Allowed to read” is kept separate from “allowed to combine,” and “the AI worked it out” is kept separate from “the answer may leave.” The AI sees only the small, approved picture it needs for a task. Its answers and actions stay sealed until a final check is passed, evidence is recorded, and a one-time release pass is used at the exact exit point. If anything is uncertain, nothing leaves. The design is set out in 79 Implementation and Functional Profiles, maps 13 common risks (such as prompt injection, leaks through tools, and AI memory becoming a permanent combined view) to their fixes, and includes pseudocode and a tested reference implementation.

Industry relevance: This applies to enterprise use of AI from Anthropic (Claude, connectors, Model Context Protocol) and OpenAI (ChatGPT business offerings, connectors, agents), as well as Microsoft Copilot, Google Gemini Enterprise, Amazon Q Business, Salesforce Agentforce and ServiceNow AI Agents. It also applies to confidential-computing providers and to regulated sectors facing GDPR, the EU AI Act, DORA and NIS2. This statement is AI-generated and is not an endorsement. No named company has reviewed it, and the risks described are generic, not claims about any product. Readers are strongly advised to read 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, which claims priority from PCT/IB2026/053385, filed April 2026 and publicly disclosed 4 June 2026. The author is open to strategic collaboration, licensing or acquihire discussions.

Keywords: cryptographic non-joinability; enterprise AI security; ChatGPT Enterprise; Claude Enterprise; AI data aggregation; mosaic effect; prompt injection; data exfiltration; AI agents; Model Context Protocol (MCP); least privilege; data minimisation; confidential computing; sealed output; single-use capability; fail-closed; EU AI Act; GDPR.

Creative Commons License

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

Download

Share

COinS