Abstract

An AI agent or operator may be authenticated, and its network, storage or security request may be approved in software. That approval is then compiled into a rule that runs repeatedly in a SmartNIC, DPU, IPU or programmable switch pipeline. The installed rule can still differ from what was approved. It may widen a wildcard, omit a mandatory rate meter, reorder priorities, or send traffic to another egress. A successful installation does not show that the hardware does only what was authorized.

Based on Indian Provisional Filing: 202631136555, filed 15 November 2026, and disclosed with the Mothership (PCT/IB2026/055615) on 16 July 2026.

This patent-pending technical concept proposes a protected software-to-hardware offload architecture. The principle is that approval of the request is not authorization of the effect the hardware will actually produce. An authorization to propose or approve an operation is kept separate from an authorization to let the installed hardware rule take effect.

How it works

  1. A protected slow path authenticates the request and records the approved decision in a protected decision record, including principal, flow, target and policy.
  2. An effect contract states the permitted match scope, transformations, state, rate limits and expiry.
  3. The rule is compiled into a proposed hardware rule for P4, Linux TC flower, DPDK rte_flow or a vendor pipeline. The materialized effect is then derived from match, priority, overlapping entries, transformations and state. Where the device supports it, the installed rule is read back instead of trusting the compiler’s intent.
  4. A correspondence check requires the effective hardware behavior to equal, or be a permitted subset of, the contract. Otherwise activation is held and the rule is re-inspected.
  5. A scoped authorization bound to the current policy and revocation epoch is issued. The device-side Finality Sink (the hardware egress, an activation latch or a destination command gate) checks it before enabling the rule. Recurring fast-path use relies on bounded local checks.
  6. After a material policy change, stale entries are fenced. Outcomes are classified as admitted, in flight, completed, denied or uncertain, and reinspection is not a blind retry.

Semantic analysis runs at rule preparation, installation and policy change, not on every packet. Existing DPDK, P4, switchdev, Open vSwitch and hardware forwarding stay in place. The architecture targets selected consequential infrastructure effects, not routine model responses.

Prospective industry relevance (AI-generated; see notice below). In each mapping, the gate sits at the real effect boundary.

  • NVIDIA (BlueField DPU, DOCA Flow): a protected control component keeps the approved decision record. It validates the effective DOCA Flow rule against the supported pipeline state. It enforces the current authorization at a protected DPU egress or activation boundary. Rules whose effective match, meter or egress exceed the approved contract are held. This depends on trusted rule readback and device authority isolation.
  • AMD (Pensando DPU, SSDK SDN policy offload): the effective P4/SDN match-action table is evaluated against the tenant’s approved decision. Current effect admission is managed on supported device hooks. Rules with a wider priority, meter, action or state are rejected. Full equivalence for arbitrary programs is not assumed.
  • Intel (Ethernet E810 eSwitch, switchdev, TC flower): an authorized tenant flow is taken before TC rule installation. The intended effect is compared with the installed match and action, within the eSwitch and driver observability available. Invalidated entry state is fenced before forwarding.
  • Amazon Web Services (Nitro System, ENA, storage offload): operation-class contracts apply to an operator-controlled resource, with enforcement at the supported Nitro-adjacent interface. Public documentation does not establish customer-accessible low-level interlocks or rule-state readback.
  • Microsoft (Azure Accelerated Networking and MANA, Copilot Studio, Azure): a customer network policy or gateway generates the protected decision record before supported NIC acceleration. A customer-deployed connector requires operation reconstruction before it calls Azure configuration controls. Protection covers only the paths the gateway routes.
  • Google Cloud (Titanium, gVNIC, agent services): operators apply admission with contract and revocation semantics, or use an external protected service gateway where Titanium internals are not accessible. This is a proposed placement, not a confirmed provider implementation.
  • OpenAI (Responses API, Agents SDK) and Anthropic (Claude tool use, Model Context Protocol): a customer-operated tool runner or MCP connector stages the candidate infrastructure change. It obtains actual rule-activation permission from an independent infrastructure controller. The Finality Sink stays at the defended network, storage or DPU effect boundary, and a tool call is not evidence of the rule hardware installed.
  • Open vSwitch and P4Runtime: known rule programming is reused. Effect-specific verification and current enforcement are added only where the hardware supports them.

These mappings identify possible complementary integration points, not missing capabilities or deficiencies in the named products. The listed vendors already provide substantial access controls, programmable policies and verification mechanisms. Practical assurances depend on actual rule readback, device authority isolation, supported enforcement interfaces, revocation semantics, enforceable path coverage and measured testing.

Where to read the full filing: search “Sangam Das” or PCT/IB2026/055615 (pctib202655615) on WIPO PATENTSCOPE (https://patentscope.wipo.int/), or read the PCT file at Technical Disclosure Commons: https://www.tdcommons.org/dpubs_series/12542/

Basis and notice: The underlying invention is human-originated and based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026. This abstract and its industry and product comparisons were prepared with AI-assisted tools and may contain errors, omissions, outdated facts or speculation. Readers must cross-check the original public PCT disclosure and the companies’ own documentation before drawing any conclusion. Company and product references are explanatory only. They imply no adoption, endorsement, validation, collaboration, licensing or infringement, and no finding on patent validity or third-party rights.

Part 10 of 39: this publication is Part 10 of a 39-part technical disclosure series, comprising 38 numbered embodiments and one separately designated Embodiment 12A.

License: The attached PDF is submitted under CC BY-NC 4.0 International, irrespective of the license displayed by the submission site.

KEYWORDS

AI Safety; Hardware Offload Security; SmartNIC; DPU; IPU; Programmable Data Plane; P4; P4Runtime; Open vSwitch; TC Flower; DPDK rte_flow; Match-Action Rule Verification; Policy-to-Hardware Equivalence; Rule Readback; Software-Defined Networking Security; Cloud Network Security; Multi-Tenant Isolation; Agentic AI Security; AI Tool-Call Security; Model Context Protocol; Consequence-Bound Authorization; Execution Finality; Finality Sink; Revocation; Fast-Path Authorization; Infrastructure Automation Security; Critical Infrastructure Protection; Zero Trust; Low-Latency Enforcement; Legacy System Compatibility

Creative Commons License

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

Share

COinS