Abstract

Enterprise AI agents are moving into sandboxed virtual machines, supervised by separate security components, connected to business systems and given persistent memory. Each layer reduces risk, but none by itself shows that a specific consequential operation, such as a database write, a message, a payment or a permission change, is still authorized at the moment it takes effect. A sandbox limits where an agent runs. A supervising monitor can detect or score suspicious behavior, and detection or risk scoring is not the same as resource-side authorization. Confidential computing protects data and code in use, and it does not by itself prevent an unauthorized downstream resource operation. Retrieved documents and memory entries can also become instructions the agent treats as authority.

Based on Indian Provisional Filing 202631125689, filed 22 October 2026, and disclosed with the Mothership (PCT/IB2026/055615), filed 4 June 2026 and publicly disclosed 16 July 2026. The Mothership is the foundational framework, and later embodiments are disclosed in the Indian filing.

This paper presents a forward-looking technical integration of that disclosure with sandboxed enterprise agent architectures. Public descriptions of Meta’s Muse architecture, announced 8 September 2026, say a separate Sentinel agent is the sole permission authority for connectors and network egress, and that real credentials never enter the agent runtime. That design shows the field is converging on separating the agent from the authority to act. This paper adds one layer: a protected enforcement domain, separate from both the agent and the supervisor, which holds authority state and releases a single-use, operation-bound authorization only after the proposed operation has been intercepted in a non-effective state, represented by a verifiable descriptor and checked against current authority. An enforcement point at the actual resource or effect boundary then re-derives the operation and refuses it if anything has changed. The supervisor remains independent of the component that permits consequential operations, so a compromised agent, a manipulated telemetry feed or a missed detection cannot by itself create authority. jigarjoshi

The paper specifies where each function needs hardware protection (key custody, attestation, protected authority state) and where an isolated software service is enough. It covers governed tool dispatch and enterprise connectors, credential brokering, revocation propagation, outstanding and asynchronous operations, cross-agent conflicts, recovery from uncertain outcomes, and persistent memory. For memory, informational content is separated from executable authority, retrieved or remembered text cannot become a privileged instruction without authorization, provenance is tracked through summaries, embeddings and derived memory, and revoked memory cannot return through stale backups, caches, synchronization or delayed jobs. Threat treatments name the trusted component, the enforcement mechanism, the remaining attack surface and the expected failure behavior for compromised agent VMs, malicious tools, stolen credentials, stale policy, connector substitution, replay, privileged-host compromise and failure of the supervisor.

The paper separates three categories: features that are publicly implemented, capabilities that have been announced, and the author’s proposed integrations. It makes no measured latency, performance or security-improvement claim. Lightweight operation-bound checks are kept apart from expensive analysis, and the approach is intended to layer onto existing agent runtimes and enterprise software without replacing them.

This is an independent, non-peer-reviewed technical proposal. It is not an endorsement, validation or infringement statement about any company or product. It does not suggest that Meta or any other company has adopted, licensed or implemented the disclosed mechanisms. Public product descriptions must be re-verified against current vendor documentation.

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

The abstract and summary in this PDF were prepared with AI assistance and may contain mistakes, omissions or outdated statements. Readers must cross-check them against the original public disclosures and the patent filing copy attached to this record before drawing any conclusion. This includes application numbers, filing and publication dates, technical descriptions, product and company references, and regulatory or legal statements. Only the original filings and official records are authoritative.

License: CC BY-NC 4.0.

Keywords

AI Safety, Trustworthy AI, Enterprise AI Agent Security, AI Agent Security, Agentic AI Security, Autonomous AI Agents, Meta Muse, Sentinel, Security Supervision for AI Agents, Independent Guardian Agent, AI Agent Sandbox, Secure Virtual Machine, Sandbox Escape Prevention, Sandbox Effect Control, Runtime Authorization, Operation-Bound Authorization, Consequence-Bound Authorization, Protected Execution-Finality, Policy Enforcement Point (PEP), Policy Decision Point (PDP), Reference Monitor, Complete Mediation, Non-Bypassable Execution Control, Zero Trust Architecture, Least Privilege, Resource-Side Authorization, Effectuation-Boundary Enforcement, Detection Versus Enforcement, Risk Scoring Limits, Tool-Call Authorization, Governed Tool Dispatch, Enterprise Connector Security, Model Context Protocol (MCP) Security, Agent-to-Agent (A2A) Security, Multi-Agent Systems Security, Delegation Chain Security, Credential Brokering, Credential Isolation, Secrets Management for AI Agents, Stolen Credential Mitigation, Connector Substitution, Egress Control, Data Exfiltration Prevention, Prompt Injection Defense, Indirect Prompt Injection, Malicious Tool Output, Manipulated Telemetry, Compromised Agent Containment, Confidential Computing, Confidential Virtual Machine, Trusted Execution Environment (TEE), Hardware Root of Trust, Hardware-Backed Key Protection, Remote Attestation, Cryptographic Operation Binding, Protected State, Privileged Host Compromise, Anti-Replay Protection, Asynchronous Retry Safety, Idempotency and Duplicate-Effect Prevention, Time-of-Check/Time-of-Use (TOCTOU) Prevention, Fail-Closed Enforcement, Revocation Propagation, Distributed Revocation, Stale Policy Mitigation, Partition-Safe Delegation, Outcome Reconciliation, Uncertain Execution Recovery, Cross-Agent Conflict Resolution, Cumulative Exposure Limits, Enterprise Database Security, Cloud Resource Control, Financial Operations Security, Agent Memory Security, Persistent Memory Poisoning, Long-Term Memory Governance, Memory Provenance, Informational Versus Procedural Memory, Derived Memory Revocation, Selective Memory Erasure, Right to Erasure, Embedding and Knowledge Graph Provenance, Stale Backup and Cache Restoration, Cross-Tenant Exposure, Context Minimization, Hardware Queue Isolation, Accelerator Security, Tamper-Evident Audit Evidence, Digital Sovereignty, EU AI Act, NIS2, Cyber Resilience Act (CRA), IETF Internet-Draft, Patent-Pending Technical Disclosure, Technical Disclosure Commons, Candidate Act, Non-Effective State, Finality Sink, Protected Enforcement Domain, Effect-Path Restriction Compilation

Creative Commons License

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

Share

COinS