--- title: [In-Depth Explanation by a Patent Attorney] Analyzing NVIDIA’s “AI Guardrail” Patent US 2024/0354319 A1 | Controlling LLMs at Runtime Without Re-training description: A patent attorney provides a thorough explanation of NVIDIA’s published application US 2024/0354319 A1 (NeMo Guardrails). Guardrails that control LLMs at runtime without retraining; the mechanism from canonical form to dialogue flow to execution; Colang; and a two-tiered approach for entry and exit image: https://www.evorix.jp/hubfs/Gemini_Generated_Image_naxriznaxriznaxr.jpg --- [Skip to content](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#main-content) - [English](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319) - [Chinese](https://www.evorix.jp/zh/blog/nvidia-llm-guardrails-patent-us20240354319) - [Deutsch (German)](https://www.evorix.jp/de/blog/nvidia-llm-guardrails-patent-us20240354319) - [Español (Spanish)](https://www.evorix.jp/es/blog/nvidia-llm-guardrails-patent-us20240354319) - [Français (French)](https://www.evorix.jp/fr/blog/nvidia-llm-guardrails-patent-us20240354319) - [Português (Portuguese)](https://www.evorix.jp/pt/blog/nvidia-llm-guardrails-patent-us20240354319) - [한국어 (Korean)](https://www.evorix.jp/ko/blog/nvidia-llm-guardrails-patent-us20240354319) - [日本語 - 日本 (Japanese - Japan)](https://www.evorix.jp/blog/nvidia-llm-guardrails-patent-us20240354319) English Show submenu for translations [![知的財産事務所エボリクス/EVORIX/大阪/特許・商標・意匠登](https://www.evorix.jp/hs-fs/hubfs/%E7%9F%A5%E7%9A%84%E8%B2%A1%E7%94%A3%E4%BA%8B%E5%8B%99%E6%89%80%E3%82%A8%E3%83%9C%E3%83%AA%E3%82%AF%E3%82%B9%20(500%20%C3%97%20250%20px).png?width=500&height=250&name=%E7%9F%A5%E7%9A%84%E8%B2%A1%E7%94%A3%E4%BA%8B%E5%8B%99%E6%89%80%E3%82%A8%E3%83%9C%E3%83%AA%E3%82%AF%E3%82%B9%20(500%20%C3%97%20250%20px).png)](https://www.evorix.jp/) - [Services](https://www.evorix.jp/en/contents) Show submenu for Services - [Patent Filing](https://www.evorix.jp/en/patent-service) - [Trademark Registration](https://www.evorix.jp/en/trademark-service) - [Design Registration](https://www.evorix.jp/en/design-service) - [Copyright](https://www.evorix.jp/en/copyright-service) - [Unfair Competition](https://www.evorix.jp/en/unfair-competition) - [Patent Attorney Advisor](https://www.evorix.jp/en/komon) - [International Filing](https://www.evorix.jp/en/international) Show submenu for International Filing - [PCT International Patent](https://www.evorix.jp/en/pct) - [Madrid Protocol Trademark](https://www.evorix.jp/en/madrid-protocol-trademark) - [Hague Design Application](https://www.evorix.jp/en/hague-design-application) - [Global Patent Systems](https://www.evorix.jp/patent-world) - [Global Trademark Systems](https://www.evorix.jp/trademark-world) - [Global Design Systems](https://www.evorix.jp/design-world) - [Industries](https://www.evorix.jp/en/industries) Show submenu for Industries - [For Startups](https://www.evorix.jp/en/startup) - [For Manufacturing](https://www.evorix.jp/en/manufacturing) - [For IT Companies](https://www.evorix.jp/it-ip-portal) - [About Us](https://www.evorix.jp/en/about) Show submenu for About Us - [Firm Overview](https://www.evorix.jp/en/about) - [Philosophy & Vision](https://www.evorix.jp/philosophy) - [Our Patent Attorneys](https://www.evorix.jp/en/staff) - [Locations](https://www.evorix.jp/en/access) - [Seminars & Lectures](https://www.evorix.jp/en/seminar) - [Careers](https://www.evorix.jp/en/recruit) - [Resources](https://www.evorix.jp/en/blog) Show submenu for Resources - [Blog](https://www.evorix.jp/en/blog) - [FAQ](https://www.evorix.jp/en/faq) - [IP Glossary](https://www.evorix.jp/en/glossary) - [Foreign Filing Subsidy Guide](https://www.evorix.jp/%E3%82%B8%E3%82%A7%E3%83%88%E3%83%AD%E5%A4%96%E5%9B%BD%E5%87%BA%E9%A1%98%E5%8A%A9%E6%88%90%E9%87%91%E3%81%AE%E3%81%94%E6%A1%88%E5%86%85) - [AI Draft × Patent Attorney Check](https://www.evorix.jp/ai-draft-patent-lp) - [Contact](https://www.evorix.jp/en/inquiry) Show submenu for Contact - [Consultation Process](https://www.evorix.jp/en/consultation) - [Inquiry Form](https://www.evorix.jp/en/inquiry) Open main navigation Close main navigation - [Services](https://www.evorix.jp/en/contents) Show submenu for Services - [Patent Filing](https://www.evorix.jp/en/patent-service) - [Trademark Registration](https://www.evorix.jp/en/trademark-service) - [Design Registration](https://www.evorix.jp/en/design-service) - [Copyright](https://www.evorix.jp/en/copyright-service) - [Unfair Competition](https://www.evorix.jp/en/unfair-competition) - [Patent Attorney Advisor](https://www.evorix.jp/en/komon) - [International Filing](https://www.evorix.jp/en/international) Show submenu for International Filing - [PCT International Patent](https://www.evorix.jp/en/pct) - [Madrid Protocol Trademark](https://www.evorix.jp/en/madrid-protocol-trademark) - [Hague Design Application](https://www.evorix.jp/en/hague-design-application) - [Global Patent Systems](https://www.evorix.jp/patent-world) - [Global Trademark Systems](https://www.evorix.jp/trademark-world) - [Global Design Systems](https://www.evorix.jp/design-world) - [Industries](https://www.evorix.jp/en/industries) Show submenu for Industries - [For Startups](https://www.evorix.jp/en/startup) - [For Manufacturing](https://www.evorix.jp/en/manufacturing) - [For IT Companies](https://www.evorix.jp/it-ip-portal) - [About Us](https://www.evorix.jp/en/about) Show submenu for About Us - [Firm Overview](https://www.evorix.jp/en/about) - [Philosophy & Vision](https://www.evorix.jp/philosophy) - [Our Patent Attorneys](https://www.evorix.jp/en/staff) - [Locations](https://www.evorix.jp/en/access) - [Seminars & Lectures](https://www.evorix.jp/en/seminar) - [Careers](https://www.evorix.jp/en/recruit) - [Resources](https://www.evorix.jp/en/blog) Show submenu for Resources - [Blog](https://www.evorix.jp/en/blog) - [FAQ](https://www.evorix.jp/en/faq) - [IP Glossary](https://www.evorix.jp/en/glossary) - [Foreign Filing Subsidy Guide](https://www.evorix.jp/%E3%82%B8%E3%82%A7%E3%83%88%E3%83%AD%E5%A4%96%E5%9B%BD%E5%87%BA%E9%A1%98%E5%8A%A9%E6%88%90%E9%87%91%E3%81%AE%E3%81%94%E6%A1%88%E5%86%85) - [AI Draft × Patent Attorney Check](https://www.evorix.jp/ai-draft-patent-lp) - [Contact](https://www.evorix.jp/en/inquiry) Show submenu for Contact - [Consultation Process](https://www.evorix.jp/en/consultation) - [Inquiry Form](https://www.evorix.jp/en/inquiry) - - [English](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319) - [Chinese](https://www.evorix.jp/zh/blog/nvidia-llm-guardrails-patent-us20240354319) - [Deutsch (German)](https://www.evorix.jp/de/blog/nvidia-llm-guardrails-patent-us20240354319) - [Español (Spanish)](https://www.evorix.jp/es/blog/nvidia-llm-guardrails-patent-us20240354319) - [Français (French)](https://www.evorix.jp/fr/blog/nvidia-llm-guardrails-patent-us20240354319) - [Português (Portuguese)](https://www.evorix.jp/pt/blog/nvidia-llm-guardrails-patent-us20240354319) - [한국어 (Korean)](https://www.evorix.jp/ko/blog/nvidia-llm-guardrails-patent-us20240354319) - [日本語 - 日本 (Japanese - Japan)](https://www.evorix.jp/blog/nvidia-llm-guardrails-patent-us20240354319) English Show submenu for translations - [Contact](https://www.evorix.jp/en/inquiry) [Contact](https://www.evorix.jp/en/inquiry) 2026/07/18 # \[In-Depth Explanation by a Patent Attorney\] Analyzing NVIDIA’s “AI Guardrail” Patent US 2024/0354319 A1 | Controlling LLMs at Runtime Without Re-training ![Picture of Takefumi Sugiura – Japanese Patent Attorney](https://app.hubspot.com/settings/avatar/4ef5aa3d824852243e3f9a0ef7c7a511) [Takefumi Sugiura – Japanese Patent Attorney](https://www.evorix.jp/en/blog/author/takefumi-sugiura-japanese-patent-attorney) Share: [facebook-f icon](http://www.facebook.com/share.php?u=https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319) [linkedin-in icon](http://www.linkedin.com/shareArticle?mini=true&url=https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319) [Twitter icon](https://twitter.com/intent/tweet?url=https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319) [pinterest-p icon](http://pinterest.com/pin/create/link/?url=https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319) [envelope icon](mailto:?body=https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319) ![Gemini_Generated_Image_naxriznaxriznaxr](https://www.evorix.jp/hs-fs/hubfs/Gemini_Generated_Image_naxriznaxriznaxr.jpg?width=1376&height=768&name=Gemini_Generated_Image_naxriznaxriznaxr.jpg) In an era where AI agents operate autonomously, one of the biggest concerns is **“runaway behavior.”** Straying from the topic, generating inappropriate content, or producing hallucinations (plausible errors)—how can we prevent these risks?NVIDIA’s answer to this challenge is patent application **US 2024/0354319 A1**, **“Runtime alignment of language models in conversational AI systems,”** which we will explore in depth in this article. In this series so far, we’ve examined patents that protect an agent’s “runtime infrastructure,” “core,” “training,” “collaboration,” and “construction.”This article addresses a new layer distinct from those—**the “safety rails (guardrails).”** It is also an interesting case in that it corresponds to “NeMo Guardrails,” which NVIDIA has released as open source. A patent attorney specializing in AI intellectual property will explain the details while citing actual claims. **💡 Key Point:** This article is part of the AI Agent Patent Series. For comparisons of various companies, please see [the Anthropic Strategy edition or](https://www.evorix.jp/blog/anthropic-patent-strategy) [the Japan, U.S., and Europe Case Studies edition](https://www.evorix.jp/blog/ai-agent-patent-cases-japan-us-europe). Table of Contents 1. [30-Second Summary | Patents Protecting AI “Safety Rails”](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#summary) 2. [Basic Information on the Patent Application](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#basic) 3. [Background | How to Prevent AI Agents from “Running Amok”](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#problem) 4. [Key Point 1 | The Concept of “Controlling at Runtime Without Retraining”](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#runtime) 5. [Key Point 2 | The Mechanism: Canonical Form → Dialogue Flow → Execution](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#mechanism) 6. [Key Point 3 | Colang | A Language for Describing AI Behavior](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#colang) 7. [Key Point 4 | A Two-Tiered Approach: Ingress and Egress](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#inout) 8. [A Clause-by-Clause Analysis of Independent Claim 1](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#claim1) 9. [Securing Intellectual Property Rights for the Agent’s “Safety Layer”](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#layer) 10. [Strategy | Balancing Open Source (NeMo Guardrails) and Patents](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#oss) 11. [How Will This Be Evaluated in Japanese, U.S., and European Examinations?](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#exam) 12. [Lessons for Our Own Applications](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#lessons) 13. [Frequently Asked Questions (FAQ)](https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319#faq) ## 30-Second Summary | Patents Protecting AI “Safety Guardrails” **● What the Patent Covers: “Guardrail”** technology **that controls and secures** LLM outputs **at runtime**. **● Key Features:** ① Controls the model at runtime **without retraining**; ② Converts user input into **canonical form (a constrained semantic representation)**; ③ Controls output by verifying (or generating new) **dialogue flows**; ④ Describes behavior using a formal language called **Colang**. **● Patent Holder:** NVIDIA Corporation. **● Status: U.S**. Patent **Application Publication (A1) = Under examination** (published October 24, 2024; 20 claims in total). **● Related Technology:** Open-source “NeMo Guardrails.” ## Basic Information on the Patent Application | Item | Content | | --- | --- | | Publication Number | US 2024/0354319 A1 | | Title of the Invention | Runtime Alignment of Language Models in Conversational AI Systems and Applications | | Publication Date | October 24, 2024 | | Filing Date / Priority Date | April 20, 2023 | | Applicant | NVIDIA Corporation | | Inventors | Razvan Dinu, Jonathan M. Cohen, Christopher M. Parisien, Traian-Eugen Rebedea | | Number of Claims | 20 (3 independent: Claims 1, 11, and 19) | | Status | Published Application (Under Examination) | | Related Technology | NeMo Guardrails / Colang | ## Background | How to Prevent AI Agents from “Running Amok” Large Language Models (LLMs) are powerful, but their outputs cannot be fully controlled. Risks such as **straying from the topic, making inappropriate statements,** or **asserting untrue information (hallucinations)** pose significant barriers to integrating AI into actual services and business operations. The impact of runaway behavior is particularly severe in AI agents that act autonomously. “**How to keep AI within safe limits**” is just as important as “**how to make AI smarter**.” This patent describes technology that solves this latter challenge. ## Core Concept 1 | The Concept of “Controlling at Runtime Without Retraining” Conventional methods for ensuring desirable AI behavior have relied on “re-training the model itself,” **such as fine-tuning or RLHF**. However, these approaches are costly and time-consuming, making them unsuitable for frequent adjustments. The approach of this patent is fundamentally different. The specification explicitly states that **“output can be controlled at runtime without training or retraining the language model.”** The core idea is to leave the model as is and place **a control layer (guardrails)** “around” it. **💡 Key Point:** The design of “controlling the model from the outside without changing it” offers the advantage of **quickly and flexibly making AI safer simply by changing settings**. Eliminating the need for the resource-intensive process of retraining—this technical benefit is a key point supporting the claim of inventive step. ## Core Concept 2 | Mechanism: Canonical Form → Interactive Flow → Execution So, how is control exercised at runtime? The process described in Claim 1 consists of the following three steps. \[Guardrail Processing Flow (Claim 1)\] ① Generation of the canonical form ・Convert user input into a “constrained semantic representation” ・Normalize diverse phrasing into a common standard form ▼ ② Determination of the Dialog Flow ・Determine the dialog flow that controls the output based on the canonical form ・“Match” against the default flow; if none exists, “generate” a new one ▼ ③ Execution of the Dialog Flow ・Execute the flow to generate the output ・= Keep the output within the boundaries of the guardrails The key point is to first convert the input **into a standardized semantic representation called the “canonical form.”** This allows diverse inputs—such as “Tell me the weather” and “What’s the weather like today?”—to be consolidated into a common format, enabling them to be matched against **a** predefined **dialog flow (i.e., a script of permitted behaviors)**. ## Core Concept ③ | Colang | A Language for Describing AI Behavior The specification explains that **a “formal (conversational or natural language) modeling language”** is used to define dialogue flows and canonical forms. This corresponds to NVIDIA’s **“Colang.”** Colang is a language that allows users to describe rules such as “If the user says this, the AI behaves that way,” as well as permitted dialogue flows (flows and subflows) and structured programming syntax. Its key feature is that it enables **the AI’s “code of conduct” to be written explicitly, much like code**. ## Core Concept ④ | A Two-Tiered Approach: Ingress and Egress According to the patent specification, the guardrails function at both **the ingress and egress points**. | Stages | Role | | --- | --- | | Ingress | Controls user input. Blocks inappropriate or dangerous input at the ingress point | | Egress | Verification of model output. Blocking undesirable output at the egress | By installing guardrails on both the “input side” and the “output side,” we **ensure safety through multiple layers**. This single framework allows us to address both defense against malicious inputs—such as prompt injection (input)—and the prevention of harmful or erroneous outputs (output). ## Reading Independent Claim 1 Clause by Clause US 2024/0354319 A1 | Claim 1 (Original Text / English) A method comprising: generating, based at least on a user input, a canonical form that comprises a constrained semantic representation of the user input; determining, based at least on the canonical form, a dialog flow that controls output of a language model; and performing one or more operations to execute the dialog flow to generate an output. Reference Translation by a Patent Attorney (Japanese) A method comprising :・a step of generating, based at least on a user input, **a canonical form** that comprises **a constrained semantic representation** of said **input**; ・a step of determining , based at least on said canonical form, a **dialog flow that controls the output of a language model**; and・a step of performing one or more operations to execute the dialog flow **to generate an output**. ### Key Points of Claim 1 Claim 1 is notable for being **quite concise and broad**, even compared to the other patents we’ve examined in this series. It is condensed into three steps: “generate a canonical form → determine a dialog flow → execute.” **💡 Key Point:** Broad claims are powerful if granted, but they **also carry a higher risk of rejection based on prior art**. This application is **pending (A1)**, and the scope within which this concise Claim 1 will ultimately be granted depends on the outcome of future examination.In practice, the standard approach is to balance risk and coverage by combining broad independent claims (1, 11, and 19) with dependent claims that include specific limitations. In addition to Claim 1 (method), Claims 11 and 19 are also filed as independent claims. ## Securing Rights for the Agent’s “Safety Layer” If we list the patents examined in this series as “components of an AI agent,” the positioning of this patent becomes clear. | Layer | Representative Patents (This Series) | What Is Protected | | --- | --- | --- | | Learning | OpenAI VPT (US 11,887,367) | Method for Learning Operations | | Main Unit/Execution | Anthropic (US 12,430,150, etc.) | Runtime and Agents | | Cooperation | OpenAI/Salesforce | Multi-Agent | | Development | OpenAI/Anthropic | Agent Creation | | **Safety (This Article)** | **NVIDIA (US 2024/0354319)** | **Guardrails and Output Control** | **💡 Key Point:** Not only “intelligence” but also **“safety and** control” are independent subjects of patent protection. As societal demands for AI governance grow, patents related to safety layers are expected to become even more important in the future. If your company possesses proprietary control and safety technologies, it is well worth considering patenting them. ## Strategy | Balancing Open Source (NeMo Guardrails) and Patents What’s interesting is that NVIDIA has **released** “NeMo Guardrails,” the technology covered by this patent, **as open source**. You might wonder, “Why file a patent if it’s open source?” but this is not a contradiction. **Balancing open source and patents:** This strategy involves widely disseminating technology **to foster an ecosystem (open source)** while maintaining **legal backing (defensive protection) through patents** for core technologies. It can be described as sophisticated IP management that achieves both widespread adoption and control.In contrast to Anthropic, which published its flagship research as a paper (without seeking a patent), NVIDIA has adopted an approach of “**publishing while securing patents**.” ## How will this be evaluated in examinations in Japan, the U.S., and Europe? ### United States (USPTO) Since Claim 1 is concise and broad, a potential point of contention is whether it will be deemed an “abstract idea (organization of information or application of rules)” under the Alice/Mayo tests. The key issue is the extent to which specific processes—such as canonical form, dialogue flow, and runtime control—will be recognized as technical implementations; the application is **currently under examination**. ### Japan (JPO) The specification describes specific data processing steps—such as converting input into canonical form and determining and executing dialogue flows—making it a structure that readily supports a claim of patent eligibility as a software-related invention. The key factors for establishing inventive step are the effect of “eliminating the need for retraining” and the specific mechanism of the guardrails. ### Europe (EPO) This configuration can be easily positioned as a technical solution to the technical problem of “safe control of LLM output” and is likely to be evaluated as a technical feature even under the COMVIK approach. A detailed comparison of examination practices for AI agent patents in Japan, the U.S., and Europe is provided in [“Patent Cases and Examination Practices in Japan, the U.S., and Europe.”](https://www.evorix.jp/blog/ai-agent-patent-cases-japan-us-europe) ## Lessons for Your Own Patent Applications **① Secure rights for “safety and control” as well.** Not only the intelligence of the AI (the model) but also the technology for controlling and ensuring the safety of its output is an important subject for patent protection. **② Highlight “no need for retraining” as a technical benefit.** The benefits of reduced costs and time serve as strong evidence of inventive step. **③ Specify unique description languages and formats.** “Mechanisms for describing behavior,” such as Colang, provide a way to move beyond abstract concepts. **④ Open source and patents can coexist.** Consider combining a dissemination strategy (OSS) with a defense strategy (patents). **⑤ Balance broad independent claims with specific dependent claims.** Since broad claims pending in court carry risks, it is important to build a multi-layered defense. Why not secure intellectual property rights for your company’s AI, including safety and control technologies? Patent attorneys with deep expertise in the IT, software, and AI fields provide comprehensive support—from claim drafting that incorporates AI safety and governance technologies, to free assessments of patentability, ensuring compatibility with open-source strategies, and filing strategies in Japan, the U.S., and Europe. [Schedule a Free Initial Consultation](https://www.evorix.jp/free-consultation)[IT & AI Intellectual Property Services](https://www.evorix.jp/it-ip-portal) ## Frequently Asked Questions (FAQ) Q. What kind of patent is US 2024/0354319 A1? A. This is a U.S. patent application by NVIDIA covering “guardrail” technology that controls and secures the output of large language models (LLMs) at “runtime.” It converts user input into a “canonical form” (a constrained semantic representation) and uses that to determine and execute an “interaction flow” that controls the output.A key feature is its ability to control the behavior of an LLM without retraining it. The application was published on October 24, 2024, and is currently under examination. It corresponds to NVIDIA’s open-source “NeMo Guardrails.” Q. What are “Guardrails”? A. It is a mechanism that “constrains and guides” AI outputs to prevent them from becoming inappropriate, harmful, or off-topic. Just as guardrails on a road prevent vehicles from veering off course, it keeps LLM outputs within a desirable range. This patent achieves this through runtime control via dialogue flow, rather than model retraining. Q. What does “control without retraining” mean? A. Conventional methods for changing AI behavior (such as fine-tuning or RLHF) require retraining the model itself, which is costly and time-consuming. The method described in this patent leaves the model as-is and instead places a control layer (guardrails) “around” it to control the output at runtime. The advantage is that safety can be enhanced quickly and flexibly simply by changing settings. Q. What is “canonical form”? A. It is a standardized “constrained semantic representation” derived from user input (e.g., summaries, definitions of intent). By normalizing diverse phrasing into a common form, it becomes easier to match the input against predefined dialogue flows. Q. Can AI safety and control technologies be patented? A. Yes, they can. As demonstrated by this patent, if specific mechanisms such as canonical form, dialogue flows, and runtime control are described, patent protection can be pursued in Japan, the United States, or Europe. AI safety and governance are fields that will become increasingly important in the future and represent a promising area for securing intellectual property rights. **Important Note Regarding This Article:** This article provides a general explanation of technology and legal systems based on published patent application bulletins. US 2024/0354319 A1 is **a published application currently under examination** and has not yet been granted. Since the claims may be amended, the final scope of protection has not yet been determined.The cited claims, abstract, and description are based on published application data (Google Patents, FreePatentsOnline, etc.); however, for legally significant purposes (FTO, infringement analysis, filing, etc.), please be sure to verify the USPTO official transcript and the latest processing information, and consult with an expert for a case-by-case review.The Japanese translation is provided for reference purposes only; the official text is the original English version. **Recommended Reading (AI Agent Patent Series):** ・[OpenAI](https://www.evorix.jp/blog/openai-multi-agent-workspace-patent-us12405822)’[s Multi-Agent Patent (Shared Workspace)](https://www.evorix.jp/blog/openai-multi-agent-workspace-patent-us12405822) ・[Anthropic](https://www.evorix.jp/blog/anthropic-computer-use-patent-us12430150)’s [Core “Computer Use” Patent (Execution Framework)](https://www.evorix.jp/blog/anthropic-computer-use-patent-us12430150) [・Patent Case Studies and](https://www.evorix.jp/blog/ai-agent-patent-cases-japan-us-europe) [Examination Practices in Japan, the U.S., and Europe (Case Studies)](https://www.evorix.jp/blog/ai-agent-patent-cases-japan-us-europe) [・Can](https://www.evorix.jp/blog/ai-agent-patent-application) [AI Agent Technology Be Patented? (Basics)](https://www.evorix.jp/blog/ai-agent-patent-application) ## Source - US 2024/0354319 A1 (Google Patents): [https://patents.google.com/patent/US20240354319A1/en](https://patents.google.com/patent/US20240354319A1/en) - Claims and Specification (FreePatentsOnline): [https://www.freepatentsonline.com/y2024/0354319.html](https://www.freepatentsonline.com/y2024/0354319.html) 特許, 外国特許, 知財戦略, AI, ソフトウェア特許, 弁理士, AI特許, 知的財産権, AIエージェント, 特許事務所, NVIDIA ## Related posts [![](https://www.evorix.jp/hs-fs/hubfs/Gemini_Generated_Image_u4oipou4oipou4oi.jpg?height=200&name=Gemini_Generated_Image_u4oipou4oipou4oi.jpg)](https://www.evorix.jp/en/blog/openai-custom-gpts-patent-us12406207) 特許, 外国特許, 知財戦略, AI, ソフトウェア特許, 弁理士, AI特許, 知的財産権, AIエージェント, 特許事務所, OpenAI ## [\[In-Depth Explanation by a Patent Attorney\] Analyzing OpenAI’s “Custom GPTs” Patent US 12,406,207 B2 | “Model Builder”: AI Creating AI](https://www.evorix.jp/en/blog/openai-custom-gpts-patent-us12406207) ![Picture of Takefumi Sugiura – Japanese Patent Attorney](https://app.hubspot.com/settings/avatar/4ef5aa3d824852243e3f9a0ef7c7a511) [Takefumi Sugiura – Japanese Patent Attorney](https://www.evorix.jp/en/blog/author/takefumi-sugiura-japanese-patent-attorney) 2026/07/18 The ability to create a unique AI assistant with specialized knowledge without writing a single... [Read more](https://www.evorix.jp/en/blog/openai-custom-gpts-patent-us12406207) [![【弁理士徹底解説】OpenAIの「動画から学ぶAIエージェント」特許 US 11,887,367 B1（VPT）を読み解く｜Computer Useの作り方](https://www.evorix.jp/hs-fs/hubfs/Gemini_Generated_Image_862ths862ths862t.jpg?height=200&name=Gemini_Generated_Image_862ths862ths862t.jpg)](https://www.evorix.jp/en/blog/openai-vpt-patent-us11887367) 特許, 外国特許, 知財戦略, AI, ソフトウェア特許, 弁理士, AI特許, 知的財産権, AIエージェント, 特許事務所, OpenAI ## [\[In-Depth Explanation by a Patent Attorney\] Decoding OpenAI’s “AI Agent That Learns from Videos” Patent US 11,887,367 B1 (VPT) | How to Create “Computer Use”](https://www.evorix.jp/en/blog/openai-vpt-patent-us11887367) ![Picture of Takefumi Sugiura – Japanese Patent Attorney](https://app.hubspot.com/settings/avatar/4ef5aa3d824852243e3f9a0ef7c7a511) [Takefumi Sugiura – Japanese Patent Attorney](https://www.evorix.jp/en/blog/author/takefumi-sugiura-japanese-patent-attorney) 2026/07/18 To “have an AI operate a computer,” you first need to teach it “how to operate it.”We’ve previously... [Read more](https://www.evorix.jp/en/blog/openai-vpt-patent-us11887367) [![【弁理士徹底解説】Anthropicの「データフロー」特許出願 US 2025/0299074 A1 を読み解く｜学習↔推論パイプラインとデータの堀](https://www.evorix.jp/hs-fs/hubfs/Gemini_Generated_Image_2hdm4x2hdm4x2hdm.jpg?height=200&name=Gemini_Generated_Image_2hdm4x2hdm4x2hdm.jpg)](https://www.evorix.jp/en/blog/anthropic-data-flow-patent-us20250299074) 特許, 外国特許, 知財戦略, AI, ソフトウェア特許, 弁理士, AI特許, マルチエージェント, 知的財産権, AIエージェント, 特許事務所, Anthropic ## [\[In-Depth Explanation by a Patent Attorney\] Analyzing Anthropic’s “Data Flow” Patent Application US 2025/0299074 A1 | Learning ↔ Inference Pipelines and the Data Trench](https://www.evorix.jp/en/blog/anthropic-data-flow-patent-us20250299074) ![Picture of Takefumi Sugiura – Japanese Patent Attorney](https://app.hubspot.com/settings/avatar/4ef5aa3d824852243e3f9a0ef7c7a511) [Takefumi Sugiura – Japanese Patent Attorney](https://www.evorix.jp/en/blog/author/takefumi-sugiura-japanese-patent-attorney) 2026/07/18 The strength of an AI agent is determined not only by the model itself, but by the entire... [Read more](https://www.evorix.jp/en/blog/anthropic-data-flow-patent-us20250299074) --- [![知的財産事務所エボリクス/EVORIX/大阪/特許・商標・意匠登](https://www.evorix.jp/hs-fs/hubfs/%E7%9F%A5%E7%9A%84%E8%B2%A1%E7%94%A3%E4%BA%8B%E5%8B%99%E6%89%80%E3%82%A8%E3%83%9C%E3%83%AA%E3%82%AF%E3%82%B9%20(500%20%C3%97%20250%20px).png?width=500&height=250&name=%E7%9F%A5%E7%9A%84%E8%B2%A1%E7%94%A3%E4%BA%8B%E5%8B%99%E6%89%80%E3%82%A8%E3%83%9C%E3%83%AA%E3%82%AF%E3%82%B9%20(500%20%C3%97%20250%20px).png "知的財産事務所エボリクス/EVORIX/大阪/特許・商標・意匠登")](https://www.evorix.jp) Intellectual Property Law Firm EVORIX   4th Floor, Qus-Hiranomachi Building, 3-1-2 Hiranomachi, Chuo-Ku, Osaka-shi, Osaka 541-0046 Japan Mail: [contact@evorix.jp](mailto:contact@evorix.jp) TEL: +81-6-7777-1884 - [Hague International Design Application](https://www.evorix.jp/hague-design-application) - [Madrid Protocol International Trademark Application](https://www.evorix.jp/madrid-protocol-trademark) - [Amazon IP Infringement Support](https://www.evorix.jp/amazon-infringement) - [About Us](https://www.evorix.jp/about) - [Our Attorneys](https://www.evorix.jp/staff) - [Services](https://www.evorix.jp/contents) - [Consultation Process](https://www.evorix.jp/consultation) - [Blog](https://www.evorix.jp/blog) - [Careers](https://www.evorix.jp/recruit) - [Trademark Portfolio](https://www.evorix.jp/trademark-list) - [Privacy Policy](https://www.evorix.jp/hubfs/%E8%B3%87%E6%96%99/HP-%E3%83%97%E3%83%A9%E3%82%A4%E3%83%90%E3%82%B7%E3%83%BC%E3%83%9D%E3%83%AA%E3%82%B7%E3%83%BC.pdf) - [Contact](https://www.evorix.jp/inquiry) ---   Copyright © 2026, Intellectual Property Law Firm EVORIX® All rights reserved.  ```json { "@context" : "https://schema.org", "@type" : "BlogPosting", "author" : { "@type" : "Person", "name" : "Takefumi Sugiura – Japanese Patent Attorney", "url" : "https://www.evorix.jp/en/blog/author/takefumi-sugiura-japanese-patent-attorney" }, "dateModified" : "2026-07-18T10:47:39.545Z", "datePublished" : "2026-07-18T10:47:39.000Z", "headline" : "[In-Depth Explanation by a Patent Attorney] Analyzing NVIDIA’s “AI Guardrail” Patent US 2024/0354319 A1 | Controlling LLMs at Runtime Without Re-training", "image" : [ "https://www.evorix.jp/hubfs/Gemini_Generated_Image_naxriznaxriznaxr.jpg" ], "mainEntityOfPage" : { "@id" : "https://www.evorix.jp/en/blog/nvidia-llm-guardrails-patent-us20240354319", "@type" : "WebPage" }, "publisher" : { "@type" : "Organization", "logo" : { "@type" : "ImageObject", "url" : "https://www.evorix.jp/hubfs/%E7%9F%A5%E7%9A%84%E8%B2%A1%E7%94%A3%E4%BA%8B%E5%8B%99%E6%89%80%E3%82%A8%E3%83%9C%E3%83%AA%E3%82%AF%E3%82%B9%20(500%20%C3%97%20250%20px).png" }, "name" : "知的財産事務所エボリクス" } } ``` ```json { "@context" : "https://schema.org", "@type" : "BlogPosting", "author" : { "@type" : "Person", "affiliation" : "知的財産事務所エボリクス", "jobTitle" : "弁理士", "name" : "杉浦 健文" }, "datePublished" : "2026-06-07", "description" : "NVIDIAの公開出願US 2024/0354319 A1（NeMo Guardrails）を弁理士が深掘り解説。LLMを再学習せず実行時に制御するガードレール、正準形→対話フロー→実行の仕組み、Colang、入口/出口の二段構え。クレーム1の原文引用、オープンソースと特許の両立戦略、日米欧の審査評価まで。", "headline" : "【弁理士徹底解説】NVIDIAの「AIガードレール」特許 US 2024/0354319 A1 を読み解く｜LLMを再学習せず実行時に制御する", "image" : "https://www.evorix.jp/hubfs/EVORIX-logo-rich.png", "inLanguage" : "ja", "mainEntityOfPage" : "https://www.evorix.jp/blog/nvidia-llm-guardrails-patent-us20240354319", "publisher" : { "@type" : "Organization", "logo" : { "@type" : "ImageObject", "url" : "https://www.evorix.jp/hubfs/EVORIX-logo-rich.png" }, "name" : "EVORIX" } } ``` ```json { "@context" : "https://schema.org", "@type" : "FAQPage", "mainEntity" : [ { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "An NVIDIA U.S. patent application covers “guardrail” technology that controls and secures the output of large language models (LLMs) at “runtime.” It converts user input into a “canonical form” (a constrained semantic representation) and uses that to determine and execute an “interaction flow” that controls the output. A key feature is the ability to control the behavior of an LLM without retraining it. The application was published on October 24, 2024, and is currently under examination. It is compatible with NVIDIA’s open-source “NeMo Guardrails.”" }, "name" : "What kind of patent is US 2024/0354319 A1?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "This is a mechanism that \"constrains and guides\" AI outputs to ensure they are not inappropriate, harmful, or off-topic. Just as guardrails on roads prevent vehicles from veering off course, this mechanism keeps LLM outputs within desirable boundaries. This patent achieves this not through model retraining, but through runtime control via the dialogue flow." }, "name" : "What is a “guardrail”?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "Conventional methods for modifying AI behavior—such as fine-tuning and RLHF—require retraining the model itself, which is costly and time-consuming. The method described in this patent controls the model’s output at runtime by placing a control layer (guardrail) “around” the model without altering the model itself. Its advantage is that it enables quick and flexible safety enhancements through simple configuration changes." }, "name" : "What does “control without retraining” mean?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "This is a conversion of user input into a standardized \"constrained semantic representation\" (e.g., summaries, definitions of intent). By normalizing input expressed in various ways into a common format, it becomes easier to match it against predefined dialogue flows." }, "name" : "What is a “canonical form”?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "That is correct. If the invention is described in terms of specific mechanisms—such as canonical form, dialogue flow, and runtime control—as in this patent, it is possible to pursue patent protection in Japan, the United States, and Europe. AI safety and governance are fields that will become increasingly important in the future and represent a promising area for patent protection." }, "name" : "Can AI safety and control technologies be patented?" } ] } ```