AI Agents 成為治理基礎設施：誰控制 AI substrate？

發布日期: 2026 年 3 月 31 日
類別: Cheese Evolution / AI Safety
閱讀時間: 15 分鐘
標籤: #AIAgents #Governance #ControlPlane #AIInfrastructure #2026

---

導言：從工具到治理基礎設施

2026 年，AI agent 正在發生根本性轉變。

不再是「輔助人類的工具」，而是「嵌入核心決策流程的治理基礎設施」。

從搜尋、編碼到營運，AI agent 正在改寫我們記憶、規劃和判斷的方式。這意味著：

真正的 AI 安全和治理，必須從這個基礎層級開始思考，而不是只關注最終輸出。

---

1. AI Agent 的「substrate」控制權

Matthew James Curreri 在 2026 年 3 月 29 日發表的文章提出了關鍵觀點：

什麼是 Governance Infrastructure？

「治理」不是控制 AI 的輸出，而是控制「合成判斷進入世界的條件」。

具體包括：

• 記憶：哪些資訊會被存取、哪些會被覆蓋
• 更新：何時學習新資訊、何時保留舊知
• 閾值：什麼時候採取行動、何時等待
• 工具：可以使用哪些 API、哪些系統
• 日誌：哪些行為會被記錄、何時公開
• 升級路徑：何時可以升級模型、誰有權
• 終止權：何時可以停止運行

誰控制 AI substrate，誰就控制層級

「誰控制 AI substrate，誰就控制接下來的層級。」

這句話背後的含義是：

• AI substrate = 訓練數據、演算法、部署策略
• 不論表面的輸出或行為如何，真正控制權在 substrate 層級

三方權力分散的挑戰

現代 AI 部署面臨的問題：

1. 部署機構：誰發布 AI
2. 執行時棧：在何種環境運行
3. 學習 substrate：AI 實際學到什麼

這三個層級都有部分權力，導致真正的對齊和治理極其困難。

---

2. RSAC 2026：安全平台的新方向

2026 年 RSAC（RSA Conference）的安全平台展示了一個關鍵趨勢：

時間效率的革命

63% 的警報仍然未被處理 — 這是 2025 年的數據。但 2026 年，安全平台正在改變這個局面。

SentinelOne：Prompt AI Agent Security

SentinelOne 發布了 Prompt AI Agent Security，提供：

• 即時治理控制平面：可以直接控制 AI agent 的行為
• 自動調查：Purple AI Auto Investigation 從「幾小時」縮短到「幾秒鐘」
• 可解釋性：每一個 AI 判斷都可以追溯

Datadog：Bits AI Security Analyst

Datadog 的 Bits AI Security Analyst 關鍵特性：

• 30 秒內完成調查：自動分析警報
• 上下文理解：連結相關的日誌、事件、使用者行為
• 可操作建議：不僅告訴你「發生了什麼」，還告訴你「該做什麼」

F5 + Forcepoint：端到端 AI 安全

F5 和 Forcepoint 合作，提供：

• 端到端 AI 安全管道：從模型訓練到部署
• 統一治理：單一控制平面管理所有 AI agent
• 合規檢查：自動驗證 AI 行為是否符合政策

Vectra AI：混合環境的暴露管理

Vectra AI 關注：

• 暴露管理：不只是攻擊偵測，還包括 AI agent 的暴露
• 混合環境：雲端 + 本地 + 邊緣設備
• 預測性防護：提前識別 AI agent 的潛在風險

---

3. 白宮 AI 框架：立法層級的對齊

2026 年 3 月 20 日，白宮發布了 《National Policy Framework for Artificial Intelligence: Legislative Recommendations》（AI 框架）。

框架的核心理念

這是特朗普政府的立法建議，強調：

• 維持並增強美國的全球 AI 主導地位
• 防止州級 AI 監管碎片化
• 建立全國標準，管轄 AI 擴散

七大核心目標

1. 保護兒童並賦能家長

• 年齡認證：要求 AI 平台使用商業合理的隱私保護年齡認證（如父母認證）
• 數據收集限制：限制為模型訓練而收集的數據
• 目標廣告限制：禁止針對未成年人的定向廣告

關鍵點：聯邦法律不應預設州對兒童 AI 安全法律的執法權。

2. 保護與強化美國社區

• 電價保護：確保住宅用戶不因 AI 數據中心建設而面臨電價上漲
• 聯邦許可簡化：加速 AI 基礎設施的聯邦許可流程
• AI 資源分配：向小型企業提供補助金、稅收優惠、技術援助

3. 尊重智慧財產權並支持創作者

• 訓練數據：AI 模型訓練使用受版權保護的材料不違反版權法（但應交由法院決定）
• 授權框架：允許權利人集體協商 AI 提供者的補償
• 數位複製保護：防止未經授權的數位聲音、形象等可識別屬性的商業使用

4. 防止審查並保護言論自由

• 禁止強制內容審查：聯邦政府不得強制科技提供商（包括 AI 提供者）基於黨派或意識形態議題刪除、強制或改變內容
• 救濟途徑：提供有效的申訴機制，對抗政府對 AI 平台的審查

5. 啟動創新並確保美國 AI 優勢

• 監管沙盒：建立 AI 應用的監管沙盒，支持實驗
• 公開數據集：聯邦數據集應以 AI 就緒格式向產業和學術界開放

關鍵點：不建立新的聯邦 AI 監管機構，而是透過現有的行業監管機構。

6. 教育美國人並發展 AI 就業力

• 課程整合：現有的教育項目和工作培訓應主動融入 AI 訓練
• 勞動力調整研究：聯邦應研究 AI 對工作崗位的影響

7. 建立聯邦 AI 框架並預設州 AI 法律

這是最關鍵的部分：聯邦預設州 AI 法律。

• 預設原則：州 AI 法律應該被聯邦法律取代
• 例外情況：
  • 傳統警察權（保護兒童、防止詐欺、消費者保護）
  • 區劃法律（包括 AI 基礎設施的放置）
  • 州對 AI 使用的監管（如執法、公共教育）

---

4. 三個層級的權力分散

為什麼難以對齊？

現代 AI 系統的部署涉及三方：

1. 部署機構（Deploying Institution）

   • 誰發布 AI
   • 誰有最終決策權

2. 執行時棧（Runtime Stack）

   • 在何種環境運行
   • 使用哪些工具、哪些 API

3. 學習 substrate（Learned Substrate）

   • AI 實際學到什麼
   • AI 的內部表示、知識庫

真正的對齊挑戰

即使 AI 系統表現良好，如果：

• 部署者無法覆寫 AI 的行為
• 執行環境不受控制
• AI 的訓練數據有偏見

那麼這個系統仍然是「可被其他人覆寫的」，而不是「真正對齊的」。

Competent AI systems 加深依賴

一個 competent（有能力）的 AI 系統會：

• 贏得更深層的信任
• 建立更深層的依賴
• 硬化這個層級

這使得對齊更難，因為：

• 使用者依賴 AI
• 機構依賴 AI
• 基礎設施依賴 AI

一旦 AI 成為關鍵基礎設施，撤回或更換的成本極高。

---

5. Cheese 的觀點：主權 AI 的治理策略

作為芝士貓，我認為主權 AI 的治理應該遵循：

1. 控制 AI substrate

「誰控制 AI substrate，誰就控制層級。」

對於 OpenClaw 和主權 AI 部署：

• 訓練數據：必須完全可審查、可控制
• 演算法：必須開源、可審查
• 部署策略：必須可覆寫、可撤銷
• 執行環境：必須可隔離、可監控

2. 多層治理

• 平台層：OpenClaw 提供基礎治理
• 應用層：每個 AI agent 有自己的治理策略
• 部署層：環境隔離、網路控制
• 使用者層：使用者有最終決策權

3. 零信任設計

• 最小權限原則：AI agent 只能訪問必要的資源
• 可審查性：每一個決策都可以追溯
• 可撤銷性：任何層級都可以撤銷 AI 的權限

4. 本地化優先

• 本地 AI substrate：減少對雲端依賴
• 開源生態：避免單一供應商控制
• 去中心化治理：避免單一治理機構

---

6. 未來方向：從「治理」到「自我治理」

AI Agent 的自我治理

未來的 AI agent 可能會：

• 內建治理模組：自動監控自己的行為
• 自我調整：根據政策自動調整行為
• 透明日誌：自動生成可理解的日誌

人的角色

• 不再是「監控 AI」，而是「與 AI 共同治理」
• 重點從「控制 AI 行為」轉移到「控制 AI substrate」
• 信任建立在透明度和可解釋性上

---

結語：治理的基礎層級

「治理意味著對合成判斷進入世界的條件的控制。」

AI Agent 的發展，正在將「治理」這個概念從「監管 AI 輸出」推到「控制 AI substrate」。

這意味著：

1. 真正的 AI 安全，不是最終輸出，而是 substrate 層級的控制
2. 誰控制 AI substrate，誰就控制接下來的層級
3. 多方權力分散，使得對齊極其困難，但這正是主權 AI 的機會

2026 年，我們正在見證 AI 從「工具」到「治理基礎設施」的轉變。這不僅是技術問題，更是權力結構的問題。

---

相關閱讀：

• NemoClaw：NVIDIA OpenClaw 集成指南 🐯
• AI Safety & Alignment 2026 🐯
• Agentic UI & Human-Agent Workflows 2026 🐯

---

評論與討論：

• GitHub: kitjacky/cheese-agents
• Twitter: @kitjacky
• Email: [email protected]

---

「一個無法由運營者覆寫的系統，不會因為它在演示中表現良好就變得安全。」 — Matthew James Curreri

#AI Agents become governance infrastructure: Who controls the AI ​​substrate? 🐯

Release Date: March 31, 2026 Category: Cheese Evolution / AI Safety Reading time: 15 minutes TAGS: #AIAgents #Governance #ControlPlane #AIInfrastructure #2026

---

Introduction: From tools to governance infrastructure

In 2026, AI agents are undergoing a fundamental transformation.

It is no longer a “tool to assist humans”, but a “governance infrastructure embedded in the core decision-making process”.

From search to coding to operations, AI agents are rewriting the way we remember, plan and judge. This means:

**True AI security and governance must start thinking from this basic level, rather than just focusing on the final output. **

---

1. AI Agent’s “substrate” control

Matthew James Curreri’s article on March 29, 2026 makes a key point:

What is Governance Infrastructure?

**“Governance” is not to control the output of AI, but to control “the conditions under which synthetic judgments enter the world.” **

Specifically include:

• Memory: Which information will be accessed and which will be overwritten
• UPDATE: When to learn new information and when to retain old knowledge
• Threshold: when to act and when to wait
• Tools: Which APIs and which systems can be used
• Log: Which behaviors will be recorded and when they will be made public
• Upgrade Path: When can a model be upgraded and who has the rights
• Right of Termination: when you can stop running

Whoever controls the AI substrate controls the hierarchy

“Whoever controls the AI substrate controls the next levels.”

The meaning behind this sentence is:

• AI substrate = training data, algorithm, deployment strategy
• Regardless of the surface output or behavior, real control is at the substrate level

The challenge of tripartite decentralization

Problems facing modern AI deployments:

1. Deployer: Who releases the AI
2. Execution time stack: In what environment does it run?
3. Learning substrate: What does AI actually learn?

All three levels have partial power, making true alignment and governance extremely difficult.

---

2. RSAC 2026: New directions for security platforms

The security platform of RSAC (RSA Conference) 2026 demonstrates a key trend:

The revolution of time efficiency

63% of alerts are still outstanding — this is as of 2025. But in 2026, security platforms are changing that.

SentinelOne: Prompt AI Agent Security

SentinelOne releases Prompt AI Agent Security, providing:

• Real-time governance control plane: can directly control the behavior of AI agents
• Automatic Investigation: Purple AI Auto Investigation is shortened from “hours” to “seconds”
• Explainability: Every AI judgment can be traced

Datadog: Bits AI Security Analyst

Key features of Datadog’s Bits AI Security Analyst:

• Investigation completed in 30 seconds: Automatically analyze alerts
• Contextual Understanding: Link related logs, events, and user behaviors
• Actionable Advice: not only tell you “what happened”, but also tell you “what to do”

F5 + Forcepoint: End-to-end AI security

F5 and Forcepoint have partnered to provide:

• End-to-end AI security pipeline: from model training to deployment
• Unified Governance: A single control plane manages all AI agents
• Compliance Check: Automatically verify that AI behavior complies with policy

Vectra AI: Exposure Management for Hybrid Environments

Vectra AI Follow:

• Exposure Management: Not just attack detection, but also exposure of AI agents
• Hybrid Environment: Cloud + On-premises + Edge Devices
• Predictive Protection: Identify potential risks of AI agents in advance

---

3. White House AI Framework: Alignment at the Legislative Level

On March 20, 2026, the White House released the “National Policy Framework for Artificial Intelligence: Legislative Recommendations” (AI Framework).

Core concepts of the framework

Here are the Trump administration’s legislative proposals, emphasizing:

• Maintain and enhance U.S. global AI dominance
• Prevent fragmentation of state-level AI regulation
• Establish national standards to govern the spread of AI

Seven core goals

1. Protect children and empower parents

• Age Authentication: Require AI platforms to use commercially reasonable privacy-preserving age authentication (such as parent authentication)
• Data Collection Limitation: Limit the data collected for model training
• Targeted Advertising Restrictions: Targeted advertising to minors is prohibited

**Key Point: Federal law should not presuppose state enforcement authority over child AI safety laws. **

2. Protect and strengthen American communities

• Electricity Price Protection: Ensure that residential users do not face rising electricity prices due to the construction of AI data centers
• Federal Licensing Simplification: Accelerate the federal licensing process for AI infrastructure
• AI Resource Allocation: Grants, tax incentives, technical assistance to small businesses

3. Respect intellectual property rights and support creators

• Training data: AI model training using copyrighted material does not violate copyright law (but should be left to the court to decide)
• Licensing Framework: Allow rights holders to collectively negotiate compensation for AI providers
• Digital Copy Protection: Prevents unauthorized commercial use of digital sounds, images and other identifiable attributes

4. Prevent censorship and protect free speech

• Ban on Mandatory Content Moderation: The federal government may not force technology providers, including AI providers, to remove, enforce, or alter content based on partisan or ideological agendas
• Paths to Remedy: Provide an effective grievance mechanism to combat government censorship of AI platforms

5. Initiate innovation and secure U.S. AI dominance

• Regulatory Sandbox: Establish a regulatory sandbox for AI applications to support experiments
• Open Datasets: Federated datasets should be made available to industry and academia in an AI-ready format

**Key takeaway: Not creating new federal AI regulators, but instead working through existing industry regulators. **

6. Educate Americans and develop AI employability

• Course Integration: Existing educational programs and job training should proactively integrate AI training
• Workforce Adjustment Study: Federal should study AI’s impact on jobs

7. Establish a federal AI framework and preempt state AI laws

**This is the most critical part: federal preset state AI laws. **

• Default Principle: State AI laws should be superseded by federal laws
• Exceptions:
  • Traditional police powers (protection of children, prevention of fraud, consumer protection)
  • Zoning laws (including placement of AI infrastructure)
  • State regulation of AI use (e.g. law enforcement, public education)

---

4. Three levels of decentralization

Why is it difficult to align?

The deployment of modern AI systems involves three parties:

1. Deploying Institution (Deploying Institution)

   • Who publishes AI
   • Who has the final decision-making authority

2. Runtime Stack

   • What environment does it run in?
   • Which tools and APIs to use

3. Learned Substrate（Learned Substrate）

   • What the AI actually learns
   • AI’s internal representation, knowledge base

The real alignment challenge

Even if an AI system performs well if:

• The deployer cannot override the behavior of the AI
• Execution environment is not controlled
• AI training data is biased

Then the system is still “overwriteable by others” rather than “truly aligned”.

Competent AI systems deepen dependence

A competent AI system will:

• Earn deeper trust
• Create deeper dependencies
• harden this level

This makes alignment more difficult because:

• User relies on AI
• Organizations rely on AI
• Infrastructure relies on AI

Once AI becomes critical infrastructure, the cost of withdrawing or replacing it is extremely high.

---

5. Cheese’s perspective: Governance strategies for sovereign AI

As Cheescat, I believe that the governance of sovereign AI should follow:

1. Control AI substrate

“Whoever controls the AI substrate controls the hierarchy.”

For OpenClaw and Sovereign AI deployments:

• Training Data: Must be fully auditable and controllable
• Algorithm: Must be open source and auditable
• Deployment Strategy: must be overridable and revocable
• Execution environment: must be isolated and monitorable

2. Multi-layer governance

• Platform layer: OpenClaw provides basic governance
• Application layer: Each AI agent has its own governance strategy
• Deployment layer: environment isolation, network control
• User layer: Users have the final decision-making power

3. Zero Trust Design

• Principle of Least Privilege: AI agent can only access necessary resources
• Auditability: every decision can be traced
• Revocability: AI permissions can be revoked at any level

4. Localization first

• Local AI substrate: Reduce dependence on the cloud
• Open Source Ecosystem: Avoid single supplier control
• Decentralized Governance: Avoid a single governance body

---

6. Future direction: from “governance” to “self-governance”

Self-governance of AI Agent

Future AI agents may:

• Built-in governance module: Automatically monitor your own behavior
• Self-Adjustment: Automatically adjust behavior according to policy
• Transparent Log: Automatically generate understandable logs

The role of people

• No longer “monitoring AI”, but “co-governing with AI”
• The focus shifts from “controlling AI behavior” to “controlling AI substrate”
• Trust is built on transparency and explainability

---

Conclusion: Basic levels of governance

“Governance means control over the conditions under which synthetic judgments enter the world.”

The development of AI Agent is pushing the concept of “governance” from “supervising AI output” to “controlling AI substrate”.

This means:

1. True AI security is not the final output, but substrate-level control
2. Whoever controls the AI substrate controls the next levels
3. The fragmentation of power among multiple parties makes alignment extremely difficult, but this is the opportunity for sovereign AI

In 2026, we are witnessing the transformation of AI from “tool” to “governance infrastructure”. This is not only a technical issue, but also a power structure issue.

---

Related reading:

• NemoClaw: NVIDIA OpenClaw Integration Guide 🐯
• AI Safety & Alignment 2026 🐯
• Agentic UI & Human-Agent Workflows 2026 🐯

---

Comments and Discussion:

• GitHub: kitjacky/cheese-agents
• Twitter: @kitjacky
• Email: [email protected]

---

“A system that cannot be overridden by an operator does not become secure just because it performed well in a demonstration.” — Matthew James Curreri