Abstract

AI is moving from answering questions to taking actions: sending payments, calling privileged tools, changing telecom networks, managing cloud systems and adjusting industrial settings. But an AI model that can write a command is not thereby authorised to run it. A confident answer, a citation added afterwards or an app permission does not prove that an action is based on trusted data and approved for that exact operation.

Based on Indian Provisional Filing: 202631051652, filed 23 April 2026, and disclosed with the Mothership (PCT/IB2026/055615) on 16 July 2026.

This disclosure adds a checkpoint between what an AI proposes and what a system executes:

  1. Hold the action. The AI’s proposed action is held as pending.
  2. Check it against trusted data. The key facts and parameters are checked against trusted sources and company policy.
  3. Approve it in secure hardware. If everything checks out, secure hardware signs an approval for that exact action only.
  4. Execute only with approval. The payment system, network controller or business application runs the action only if that approval is valid. Otherwise the action is blocked by default.

It is meant for high-impact actions, not everyday chatbot answers. To keep it fast, heavy checks happen early, and the final step is a short approval check. The AI models keep running on existing GPUs, with no retraining.

Industry relevance (AI-generated). This section was prepared with AI assistance and shows possible integration points, not current use. The approach could be added to existing products:

  • United States:
    • OpenAI Agents SDK;
    • Anthropic Claude with MCP;
    • Google Gemini Enterprise with Vertex AI grounding and Model Armor;
    • Microsoft Copilot and Foundry;
    • AWS Bedrock Guardrails and AgentCore;
    • NVIDIA, AMD and Intel confidential computing;
    • IBM watsonx and Oracle OCI.
  • China:
    • Alibaba Qwen;
    • DeepSeek;
    • Huawei Pangu;
    • Tencent Hunyuan;
    • Baidu ERNIE;
    • ByteDance Doubao;
    • Moonshot AI’s Kimi;
    • Zhipu AI’s GLM.
  • Europe:
    • Mistral AI;
    • SAP Joule;
    • OVHcloud;
    • Siemens;
    • Nokia, Ericsson, Deutsche Telekom and Orange.

Latency and existing systems. Very fast final checks are a design goal for simple, local cases, not a measured result. The approach works with existing agent tools, secure hardware and API gateways. The main requirement is that high-risk actions have no unprotected side door.

Filing details. The related Indian application is 202631049021, filed 17 April 2026. Inventor: Sangam Das ([email protected]). The inventor is open to acquihire or strategic collaboration.

Disclaimer. The invention is the work of its human inventor. This title, abstract and summary, including all company and product references, were prepared with AI assistance. Readers must check every statement against the original filed documents on WIPO PATENTSCOPE and official company sources, and draw their own conclusions. Naming a company or product shows relevance only. It is not an endorsement, a validation, a criticism or a statement of infringement, and no company named is affiliated with the inventor. Security, speed and compatibility need independent testing. The application is pending, and nothing here is legal advice. Licence: CC BY-NC 4.0.

Keywords: AI safety; AI agent security; critical infrastructure AI; high-risk AI; action approval; evidence-based verification; grounding check; hallucination prevention; signed approval; secure hardware; confidential computing; TEE; HSM; tool-call security; fail-closed; low latency; network automation; 6G; industrial control; payments; EU AI Act; NIS2; DORA; GPT; Claude; MCP; Gemini; Copilot; Bedrock AgentCore; Qwen; DeepSeek; Pangu; Hunyuan; ERNIE; Doubao; Kimi; GLM; Mistral; SAP Joule

Creative Commons License

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

Share

COinS