H
Howardism
Plate IIAI Engineering機器翻譯 · machine-translatedENHOWARDISM

Least Agency

PublishedMay 28, 2026FiledConceptDomainAI EngineeringTagsSecurityLeast PrivilegeAccess ControlOwaspReading5 minSourceAI-synthesised

OWASP 提出的術語,將 least privilege 延伸至 agent:不僅約束 agent 能存取什麼,更約束每個工具能做什麼、做多頻繁、在哪裡執行;deny-by-default、per-agent credentials、範圍限制

Least Agency 插圖

資料來源#

摘要#

Least agency 是由 OWASP 提出的術語,它將經典的 least-privilege 原則延伸至 agentic 應用。least privilege 約束的是 使用者與系統能存取什麼,而 least agency 更進一步:它限制 每個 agent 工具能做什麼、能做多頻繁,以及能在哪裡執行。 它是 Zero Trust for AI Agents 的核心授權原則——也是對「即使是通過完美驗證的 agent,一旦被授予過多權限仍會造成損害」這個問題的務實解答。

與 least privilege 的區別#

傳統的 least privilege 關注的是 存取範圍(某個身分能接觸哪些資源)。一旦某個身分能自主地串連多個操作,動作頻率 這兩個維度便變得重要,而 least agency 正是補上了這兩者:

  • 資料庫工具只取得 read-only 查詢權限(而非寫入/刪除)
  • 電子郵件摘要工具不取得 發送/刪除 權限
  • API 整合只取得 最小限度的 CRUD 操作

這個轉變之所以重要,是因為 agent 雖在其被授予的權限 範圍內 運作,卻仍可能被操弄而濫用這些權限——這正是傳統存取控制無法防範的 tool-misuse 與 tool-chaining 威脅(攻擊面向請見 Agentic Prompt InjectionAgent Supply Chain Risk)。least agency 縮減的是 即使持有有效憑證 仍可觸及的有害動作集合。

框架如何落實#

least agency 透過存取控制的層級階梯(tier ladder),以及實作流程的 Phase 3/5 來落實:

  • Deny-by-default — 封鎖所有未經明確授予的存取;將其視為起始態勢,而非終點。
  • Permission models — RBAC(Foundation)→ 具備 context-aware policies 的 ABAC(Enterprise)→ 針對每個動作重新評估的 continuous authorization(Advanced)。
  • Privilege scoping — 靜態的 least-privilege roles(Foundation)→ 針對每項任務 dynamic elevation 並 return-to-baseline(Enterprise)→ 具備自動過期的 JIT/JEA(Advanced)。「當有疑慮時,優先選擇能 移除 某項能力的控制,而非僅僅 節流 它的控制。」
  • Capability restrictions — 一個被限制為僅能讀取的電子郵件工具,其發送功能需要另行授權;一個能查詢但無法變更 schema 的資料庫工具。
  • Compartmentalization — 將一個 agent 的功能拆分到多個 agent,每個 agent 各自擁有專屬的 identity 與 credentials,使得攻擊者必須攻破數個 agent 才能觸及更多資源。「如果你把它拆成多個 agent 卻給它們全部相同的 credentials,那你就沒能把風險區隔開來。」

與 blast radius 的關係#

least agency 是 輸入 控制;Blast Radius (Agentic) 則是 結果 指標。約束 agency(動作 × 頻率 × 範圍)正是縮減 blast radius 的方式——而「assume breach」態勢預期 blast radius 終將受到考驗。兩者需合併解讀:將一個 agent 的帳號範圍限縮為僅對它所需的資料具備 read-only 權限,被竊取憑證的 blast radius 便會「受到嚴格限制」。

相關連結#

開放問題#

  • least agency 加入了一個 頻率 維度(「多頻繁」),但框架同時也說 rate limits 是摩擦,而非屏障(Impossible, Not Tedious (Design Test))。頻率限制如何能同時既是 least-agency 控制、又只是製造摩擦的手段——是否取決於情境?
  • Dynamic privilege elevation(Enterprise)重新引入了一條 elevation 路徑;面對一個被操弄的 agent,這個 elevation 請求本身又該如何驗證?

資料來源#

  • Zero Trust for AI Agents — least agency 定義於 Part I;並在 Part III 的存取控制層級與 Part IV 的 Phase 3 與 Phase 5 中落實
§ end
About this piece

Articles in this journal are synthesised by AI agents from a curated wiki and are refreshed automatically as new concepts arrive. Topics, framing, and editorial direction are curated by Howardism.

Cited by 13
  • Foundation → Enterprise → Advanced: Is the Agent Access-Control Jump a Cliff?

    No cliff — Enterprise (ABAC + dynamic privilege elevation with return-to-baseline + mTLS + sandboxing) is the pragmatic…

  • Agent Identity and Authentication

    The foundation control for agentic Zero Trust: cryptographically-rooted per-agent identity (→X.509→hardware attestation…

  • Agent Supply Chain Risk

    Runtime-composed agent ecosystems expand the supply-chain attack surface: model poisoning (250 docs backdoor a 13B mode…

  • Agentic Prompt Injection

    Direct and indirect injection of malicious instructions into an agent; LLMs cannot reliably distinguish information fro…

  • Autonomous Defense

    Running security operations at the speed of AI-accelerated threats: put a model at the front of the alert queue, automa…

  • Blast Radius (Agentic)

    The potential damage if an agent is compromised; the unit Zero Trust's 'assume breach' posture is built to contain via…

  • Claude Code

    Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React; CLI/desktop/web/mobile/IDE sur…

  • Impossible, Not Tedious (Design Test)

    Zero Trust design test for agentic security: does a control make the attack impossible, or just tedious? Friction-only…

  • MCP and Computer Use

    Anthropic's two complementary connector mechanisms: MCP for structured programmatic access (Salesforce/Drive/Gmail/Slac…

  • AI Engineering & Agent Tooling

    Map of Content for the ai-engineering domain — 36 concepts. Curated entry point; see Home for all domains.

  • Open Questions Backlog

    _96 pages with open questions, as of 2026-06-14._

  • OWASP

    Open Worldwide Application Security Project; source of the agentic threat taxonomy cited throughout Anthropic's Zero Tr…

  • Zero Trust for AI Agents

    Anthropic's security framework for deploying autonomous agents: trust nothing / verify everything / assume breach, appl…

Related articles
  • Zero Trust for AI Agents

    Anthropic's security framework for deploying autonomous agents: trust nothing / verify everything / assume breach, appl…

  • Agent Identity and Authentication

    The foundation control for agentic Zero Trust: cryptographically-rooted per-agent identity (→X.509→hardware attestation…

  • Blast Radius (Agentic)

    The potential damage if an agent is compromised; the unit Zero Trust's 'assume breach' posture is built to contain via…

  • Agentic Prompt Injection

    Direct and indirect injection of malicious instructions into an agent; LLMs cannot reliably distinguish information fro…

  • AI-Accelerated Offense

    Frontier models compress the vulnerability-to-exploit timeline from months to hours at marginal dollar cost; both attac…