Abstract
Did an independent WIPO patent filing anticipate Meta's new Muse/Sentinel AI agent security architecture by two months? Here is a technical teardown
This disclosure describes an architecture for securing autonomous AI agents. The agent runs inside a sandbox that holds no real credentials. Every consequential action, including outbound network requests, API calls, and credential use, must pass a separately enforced authorization gate on the host before it takes effect. The agent may propose an action, but the authority to execute it sits outside the sandbox.
The architecture was published as Embodiment 85.B2, “Agent Sandbox Egress Finality Gate,” in international application PCT/IB2026/055615, published by the World Intellectual Property Organization in July 2026. Its elements include:
- separation of action proposal from action authorization
- host-side enforcement at the sandbox boundary
- provenance and taint handling for data entering the agent’s context
- control of credential use, API calls, and network egress outside the agent’s reach
On 8 September 2026, Meta publicly described Muse/Sentinel, an AI-agent security design that also uses a credential-less sandbox and a protected host component that decides whether outbound requests proceed.
This record maps the components of the two designs to each other. Each correspondence is classified as one of four types:
- stated expressly in the earlier publication
- a necessary consequence of what is stated
- a reasonable technical mapping
- not supported by the documents
The goal is to separate the shared architectural principle from implementation choices specific to Meta, such as container runtime, credential broker, proxy, policy engine, and hardware placement. This lets examiners and practitioners assess the earlier publication when evaluating later filings in this field.
The full mapping and review package are at https://zenodo.org/records/23255613.
Keywords / hashtags
#AIAgentSecurity #AgentSandbox #SandboxEgressControl #CredentialIsolation #CredentiallessSandbox #HostEnforcedAuthorization #ProposeVsAuthorize #TaintTracking #DataProvenance #EgressProxy #AgenticAI #MetaMuse #MetaSentinel #PriorArt #DefensivePublication #PCT #WIPO #ExecutionFinality
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License
Recommended Citation
Das, Sangam Kumar, "Host-Enforced Egress Authorization for Credential-Less Sandboxed AI Agents: Correspondence Between Embodiment 85.B2 of PCT/IB2026/055615 (WIPO, June 2026) and Meta Muse/Sentinel (September 2026)", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12603