Abstract

AI Agents Can Be Authorized to Act — But Who Proves the Real Action Happened? The Missing Execution-Security Layer for Google, OpenAI and Anthropic

Technical Subtitle: From Google Gemini and Enterprise Agent Platforms to OpenAI Responses API and Anthropic Claude — How Receipt-Gated Execution, Authenticated Real-World Evidence, Independent Validation and Protected Continuation Authority Could Strengthen the Security of Autonomous AI Agents.

ABSTRACT

Artificial intelligence is undergoing a fundamental transition from systems that generate responses to autonomous agents capable of producing consequential real-world effects. Modern AI agents can invoke external tools, modify production databases, deploy software, change cloud configurations, initiate financial transactions, and release sensitive enterprise information.

Technology companies including Google, OpenAI and Anthropic provide agent-development frameworks, tool interfaces, identity and authorization mechanisms, approval workflows, and security controls intended to support increasingly capable autonomous systems.

However, a critical distinction remains between authorizing an AI agent to perform an operation and independently verifying that an earlier real operation actually produced the approved outcome before permitting a larger consequential action.

An agent may possess valid credentials, satisfy policy checks, and receive permission to execute a tool call while the actual destination, transaction state, deployed resource, or resulting effect differs from what was authorized.

A tool-generated success response, execution log, policy approval, or simulated result does not necessarily prove that the intended operation was completed at the approved destination.

Executing the entire operation merely to establish whether the real execution path works may expose organizations to irreversible data disclosure, production disruption, unauthorized financial consequences, or other operational risks.

This technical study examines a proposed Receipt-Gated Interim Effectuation Validation (IEV) architecture and its potential application to three major AI-agent ecosystems: Google Gemini, OpenAI and Anthropic Claude.

The architecture introduces a protected execution sequence in which an authorized agent first performs a small, real, strictly bounded operation through an approved execution path.

The actual destination or a protected independent observer then produces an authenticated Effect-Confirmation Receipt (ECR) identifying the operation, destination, observed effect, execution phase, and relevant authorization conditions.

An independent validator, operating outside the agent's control, verifies that evidence against the approved operation, destination, permitted scope, expected effect, freshness requirements, and applicable policy or revocation conditions.

Only after successful verification does a separately protected enforcement component permit the broader consequential operation.

The proposed technical distinction is not another allow-or-deny filter for AI tool calls. It is the conditional release of subsequent execution authority based on independently validated evidence that an earlier bounded real-world effect actually occurred as authorized.

Where a receipt or acknowledgement is missing, the system does not automatically assume that the operation failed. Instead, it enters a protected indeterminate-outcome state, seeks authoritative destination records, and prevents unsafe repeated or expanded execution until reconciliation requirements are satisfied.

The study examines three prospective implementation pathways:

  • Google: Protected tool execution through Gemini-related enterprise agent infrastructure, Agent Development Kit, agent gateways, Apigee, and Cloud Run, with independent verification of bounded cloud or business operations before broader execution.

  • OpenAI: Developer-controlled execution boundaries around Responses API, Agents SDK, function calling, and MCP integrations, where authenticated database commits or destination-state evidence govern subsequent tool operations.

  • Anthropic: Protected client-side tools, Claude Agent SDK integrations, Claude Code workflows, and MCP services, where authenticated deployment or transaction evidence must be validated before authorizing a larger operation.

The proposed architecture is designed to complement existing identity, policy enforcement, human approval, model security, observability, and isolation controls rather than replace them.

The study also examines alternate execution routes, independent authority separation, forged success messages, stale or replayed receipts, lost acknowledgements, latency overhead, legacy compatibility, and the distinction between software-controlled tool execution and provider-managed operations requiring deeper enforcement integration.

The central research objective is to determine whether AI-agent security can progress beyond verifying who is permitted to act toward independently proving what actually happened before allowing the next consequential action.

These company-specific mappings represent proposed engineering integrations and require independent verification, implementation, and testing.

Part 4 of 8 — Frontier AI Agents, Tool Execution and Evidence-Gated Continuation

This is Part 4 of an eight-part technical disclosure series.

Part 1 introduces the foundational architecture and explains the central concept for general readers.

Part 2 examines prospective applications involving Huawei telecommunications systems, Mastercard agentic payment infrastructure, and OpenAI agent-execution interfaces.

Part 3 explores enterprise AI, cloud execution, and accelerated computing through Microsoft, AWS, and NVIDIA.

Part 4 focuses on three major frontier AI ecosystems — Google, OpenAI and Anthropic — examining how independently validated real-world execution evidence could complement their respective agent-development and tool-integration environments.

The emphasis is on practical engineering insertion points, not hypothetical changes to AI model weights or assumptions that existing provider-managed execution environments can automatically be controlled by an external application.

The Core Idea in Plain Language

Before allowing a large or potentially irreversible operation, the proposed system verifies a smaller, authorized, real operation through the actual destination or a protected observation mechanism.

The sequence is straightforward:

  1. Perform a small real action. An AI agent initiates a narrowly authorized operation through a protected tool service or execution boundary.

  2. Collect authenticated evidence. The destination system or an independent protected observer generates a signed or otherwise cryptographically authenticated receipt describing the actual operation and observed outcome.

  3. Verify the evidence independently. A validator outside the agent's control checks the receipt against the original authorization, destination, operation identity, permitted scope, expected effect, and current security conditions.

  4. Release subsequent execution authority. Only when the receipt satisfies the required validation conditions does a protected enforcement component authorize the broader action.

No valid qualifying receipt means no release of the larger consequential operation.

If the receipt is lost or the outcome remains uncertain, the system checks authoritative destination records rather than assuming failure and blindly retrying.

The initial bounded action does not guarantee future success. The larger operation remains subject to fresh authorization, policy continuity, destination binding, and protected enforcement at its own execution boundary.

Industry Alignment 1 — Google: Gemini, Enterprise Agents and Protected Cloud Operations

Relevant Technology Directions: Gemini, Gemini Enterprise Agent Platform, Agent Development Kit (ADK), Agent Gateway, Agent Identity, Agent Registry, Model Armor, Apigee, Cloud Run, and Cloud KMS.

Google's enterprise AI and cloud infrastructure provide potential environments for studying how receipt-gated execution could complement existing agent identity, gateway security, application governance, and model-protection capabilities.

Proposed Technical Integration

An enterprise AI agent proposes modifying cloud resources, updating business records, or executing a consequential workflow.

A protected tool service, potentially deployed behind a supported gateway or API-management interface, intercepts the proposed operation before broader execution.

The protected service first performs a narrowly scoped real action, such as updating one approved resource, committing a limited database transaction, or applying a controlled configuration to a small production cohort.

The destination system generates authenticated state evidence identifying what actually occurred.

An independent validator compares that evidence with the approved operation, permitted destination, resource identity, scope, and relevant policy conditions.

Only when the result satisfies the required conditions does the protected execution service authorize the broader operation.

Existing gateway policies, agent identity controls, application permissions, and Model Armor protections remain complementary security layers.

The distinguishing mechanism is not simply whether the agent is permitted to invoke a tool. It is whether independently verified destination evidence permits the next consequential execution phase.

Implementation requires an enforceable developer-controlled or destination-side boundary; an ordinary gateway policy is not assumed to prove external effectuation by itself.

Industry Alignment 2 — OpenAI: Responses API, Tool Execution and Verified Database Commit

Relevant Technology Directions: OpenAI Responses API, Agents SDK, Function Calling, Model Context Protocol (MCP), and supported human-in-the-loop approval workflows.

OpenAI's agent-development interfaces offer a practical environment for investigating execution-evidence controls when the application developer owns the consequential tool executor.

Proposed Technical Integration

An AI agent proposes updating a large number of enterprise records or invoking a consequential external API.

The application-controlled tool executor does not immediately release the entire operation.

Instead, it performs a small, approved real change against a strictly bounded record set or controlled destination.

The protected destination returns authoritative transaction-state evidence, such as a verified database commit record or an authenticated result from an independent destination observer.

A validator checks whether the actual committed operation matches the approved destination, record scope, transaction identity, and policy requirements.

Only after successful verification does a protected executor permit the wider update.

A tool response reporting “success” is not sufficient evidence that the intended database commit actually occurred.

Practical Enforcement Boundary

Where the application controls the tool function, gateway, and destination credentials, the architecture can be prototyped without modifying the underlying AI model.

However, provider-managed tools and execution paths that are not controlled by the application may require destination-side enforcement, a supported approval interface, or additional provider cooperation.

An application-level wrapper cannot be assumed to govern every external operation performed through an independently managed route.

This distinction is essential when evaluating actual bypass resistance.

Industry Alignment 3 — Anthropic: Claude, MCP and Verified Agent-Driven Deployments

Relevant Technology Directions: Claude API, Claude Agent SDK, Claude Code, Model Context Protocol (MCP), and supported tool hooks or client-side execution controls.

Anthropic's agent-development ecosystem provides another potential integration environment for independently validated, consequence-bound operations.

Proposed Technical Integration

A Claude-based agent proposes deploying software, transferring sensitive enterprise information, or executing a privileged external operation.

A protected client-side tool executor or MCP service first permits a small, strictly authorized action.

For example, a deployment service may apply a canary release to one approved resource or a controlled group of instances.

The deployment destination generates authenticated evidence of the actual deployed version, resource identity, operational state, and relevant release conditions.

A separate validator verifies that the observed deployment matches the approved candidate operation.

Only after successful verification does the protected deployment controller authorize wider rollout.

The same general mechanism could be adapted to database updates, privileged workflows, or approved data transfers, provided the system can enforce bounded consequences and verify appropriate destination evidence.

The proposed security boundary resides at the protected execution service, not merely in the agent's generated instructions or self-reported tool results.

A local hook, including a PreToolUse-style interception mechanism, may support application-side control but must not be assumed to govern every hosted, managed, delegated, or alternate execution route.

Broader enforcement requires appropriate control over destination credentials, tool gateways, protected resource interfaces, or other supported execution boundaries.

What This Technical Study Also Covers

The publication evaluates the architectural and experimental requirements for deploying receipt-gated execution across different AI-agent ecosystems.

Execution Insertion Points: Developer-controlled function executors, MCP gateways, protected tool services, cloud API boundaries, deployment controllers, and destination-side enforcement interfaces.

Identity and Authority Separation: Distinguishing agent identity, delegated tool authority, independent validator authority, and protected continuation authorization.

Authenticated Evidence: Destination-generated receipts, authoritative transaction records, signed deployment outcomes, verified resource states, freshness information, operation identifiers, and anti-replay controls.

Protected Continuation State: Maintaining release authority outside the agent's control, preventing unauthorized modifications, stale approvals, or attempts to continue through alternate routes.

Threat Testing: Alternate credentials, forged success messages, replayed evidence, destination substitution, revoked authority, partial execution, lost confirmations, and bypass attempts.

Latency-Aware Deployment: Performing receipt validation at consequential execution boundaries rather than unnecessarily gating individual model tokens, routine reasoning steps, or low-risk internal computation.

Legacy Compatibility: Adding protected services or enforcement controls around existing tools, APIs, databases, and deployment systems without assuming modifications to foundation-model weights.

Software-First Versus Deeper Integration: Identifying which mechanisms can be demonstrated within developer-owned infrastructure and which require provider cooperation, supported managed-service controls, hardware isolation, or destination-side enforcement.

The technical evaluation distinguishes architectural feasibility from demonstrated security and performance results. Claims of non-bypassability, low latency, interoperability, or verified external effectuation require independent experimental evidence.

Foundational Patent and Technical References

Mothership International Patent Application: PCT/IB2026/055615 Filed: 4 June 2026 Publicly disclosed: 16 July 2026

Related Indian Provisional Patent Application: 202631117633 Filed: 1 October 2026

Related IETF Internet-Draft: Evidence Is the Key, Not the Log — Receipt-Gated Interim Effectuation Validation

Internet-Draft Identifier: draft-das-receipt-gated-iev

Read the IETF Internet-Draft

The inventor identifies 1 October 2026 as an initial disclosure date associated with the specific receipt-gated formulation. The exact public availability date of the IETF Internet-Draft and any patent-priority implications should be independently verified from the relevant records.

Patent-Pending Technical Concept — DAS Protocols

Non-Endorsement, Independent Verification and AI-Assistance Notice

Company and product names are included exclusively to illustrate prospective technical integration opportunities and potential industry relevance.

Nothing in this publication states or implies that Google, OpenAI, or Anthropic has implemented, adopted, licensed, endorsed, partnered on, or independently validated the proposed invention.

The publication does not allege copying, patent infringement, technical deficiencies, security failures, or non-compliance by any named company.

The company-specific industry-alignment analysis was prepared with AI assistance and may contain assumptions, inaccuracies, or descriptions requiring independent verification.

Readers should cross-check all proposed integrations against the original patent disclosures, current official vendor documentation, supported platform interfaces, applicable technical standards, and reproducible experimental results before drawing legal, technical, or commercial conclusions.

Access to the Foundational PCT Filing

The foundational patent materials may be located through WIPO PATENTSCOPE using:

Inventor: Sangam Das PCT Application Number: PCT/IB2026/055615 Alternative Search: pctib202655615

Additional technical publication:

TD Commons — Foundational PCT Reference

Strategic Collaboration and Licensing

The inventor is open to strategic technical collaboration, technology licensing discussions, and acquihire opportunities involving frontier AI developers, enterprise agent platforms, cloud infrastructure, confidential computing, and critical infrastructure security.

Publication License: Creative Commons Attribution–NonCommercial 4.0 International (CC BY-NC 4.0).

The publication licence does not grant patent rights or authorization to practise patent-protected subject matter. Commercial implementation requires a separate written licence from the relevant rights holder.

Keywords and Technical Search Index

Core Architecture: Receipt-Gated Execution; Receipt-Gated IEV; Interim Effectuation Validation; Effect-Confirmation Receipt (ECR); Staged Execution; Bounded Real-Effect Validation; Canary Deployment; Progressive Rollout; Conditional Commit; Signed Receipt; Authenticated Outcome Evidence; Independent Validation; Protected Continuation Authority; Cryptographic Release Authorization; Protected Execution-Finality.

Security and Enforcement: AI Agent Security; Agentic AI; Zero Trust; Policy Enforcement Point; Reference Monitor; TOCTOU Prevention; Destination Binding; Replay Prevention; Independent Runtime Validation; Authenticated Transaction State; Fail-Closed Enforcement; Protected Tool Invocation; Non-Bypassable Execution; Authority Separation; Indeterminate Outcome Reconciliation; Human-in-the-Loop Approval.

Google Technologies: Google; Google Gemini; Gemini Enterprise Agent Platform; Agent Development Kit; ADK; Agent Gateway; Agent Identity; Agent Registry; Model Armor; Google Cloud; Apigee; Cloud Run; Cloud KMS; AI Agent Governance; Cloud Resource Validation; Enterprise Agent Security.

OpenAI Technologies: OpenAI; Responses API; Agents SDK; Function Calling; Tool Execution; MCP Tools; Model Context Protocol; Developer-Controlled Tool Executor; Human-in-the-Loop Approval; Verified Database Commit; Enterprise AI Agents; Agent Tool Security; Authenticated External Effects.

Anthropic Technologies: Anthropic; Claude; Claude API; Claude Agent SDK; Claude Code; Model Context Protocol; MCP; PreToolUse Hook; Protected MCP Server; Canary Deployment; Verified Deployment Evidence; Autonomous Software Engineering; Agentic Workflow Security; Protected Data Transfer.

Patent and Standards References: DAS Protocols; Patent Pending; PCT/IB2026/055615; Indian Provisional Application 202631117633; Evidence Is the Key, Not the Log; draft-das-receipt-gated-iev.

Creative Commons License

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

Share

COinS