資料來源#
摘要#
Zero Trust for AI Agents 套用到每一項控制措施的一個設計審查問題:這會讓攻擊變得不可能,還是只變得繁瑣? 價值來自摩擦而非硬性屏障的控制措施——額外的樞紐跳轉、速率限制、非標準連接埠、以 SMS 為基礎的 MFA——面對能大規模反覆執行繁瑣步驟的對手時,效力會急遽下降。這個框架之所以重要,是因為代理式攻擊者擁有無限耐心,且每次嘗試的成本近乎為零:「這會花太久時間,不值得做」這類內建於人類判斷中的假設,已不再成立。
能存續的控制模式#
通過測試的控制措施具有一項結構性特徵——它們是移除能力,而非限流:
- 硬體綁定的憑證(無法被竊取,不只是難以竊取)
- 會過期/短生命週期的權杖(視窗會關閉,不只是變窄)
- 密碼編譯身分(偽造在計算上很困難,而不只是帶來不便)
- 不存在的網路路徑,而非僅僅不方便的路徑
這個框架的經驗法則是:「有疑慮時,優先選擇移除能力的控制措施,而不是限制能力的控制措施。」
應用場景#
這項測試會影響每一層的建議,並明確出現在各個決策點:
- 提高基礎門檻——僅靠摩擦的控制措施(可從 lockfile 中以 grep 找出的輪換式長效 API 金鑰、SMS MFA、速率限制)即使在入門層級也不再合格。
- 爆炸半徑評估(Phase 3)——「如果你的遏制計畫依賴摩擦……就假設它會失敗。」請參閱 Blast Radius (Agentic)。
- 工具沙箱化(Phase 5)——「速率限制是摩擦,不是屏障:它們能爭取時間,卻無法阻止堅定的代理式攻擊者。」
脈絡與收斂#
這與 LLM-Driven Vulnerability Research 中獨立提出的論點相同:該文指出,「價值來自讓利用漏洞變得繁瑣的緩解措施,面對能以低成本反覆執行繁瑣步驟、由模型輔助的對手時會變弱」,而硬性屏障(KASLR、W^X)仍然很重要。兩個來源得出相同結論:攻擊性研究以實證發現摩擦會失效;安全框架則將這項發現轉化為規範性的設計測試。兩者都源自 AI-Accelerated Offense——每次嘗試的成本近乎為零,正是 AI 加速帶給攻擊者的能力。
Out-of-Band Prompt-Injection Defense 文獻是第三個、更鮮明的例子。其核心做法——透過確定性參考監視器,移除低完整性主體授權高完整性動作的能力,而不是限制這項能力——正是將這項測試應用於提示注入防禦。Narisetty 等人(2026)的評估也以實證提供了兩個部分:自適應攻擊讓in-band(類似摩擦、模型內部)防禦的成功率升至 >90%,而移除能力的out-of-band閘門,在一項手工設計的自適應攻擊下仍然守住防線。「不可能,而非繁瑣」正好預測了這種不對稱性。
相關連結#
- Zero Trust for AI Agents — 將此採納為固定設計審查問題的框架(樞紐)
- AI-Accelerated Offense — 說明每次嘗試的成本為何趨近於零,也就是摩擦控制措施失效的原因
- Blast Radius (Agentic) — 依賴摩擦的遏制計畫會無法通過測試
- LLM-Driven Vulnerability Research — 對相同摩擦退化發現的獨立實證陳述
- Least Agency — 「移除能力而非限制能力」是以最小代理性表述的啟發式原則
- Out-of-Band Prompt-Injection Defense — 確定性參考監視器會移除低完整性資料授權高權限動作的能力;論文中 in-band 失效/out-of-band 維持有效的結果,正是這項測試預測的不對稱性,且已獲得測量
- Agent Data Injection (ADI) — 清楚呈現不可能與繁瑣的對比:每次操作的使用者確認對話框是會失效的摩擦控制措施(代理的推理會強化攻擊者偽造的故事),而元素 ID 的nonce 隨機化(ChatGPT Atlas)移除了預測識別碼的能力,並維持有效
- Task-Specification Effects in Prompt Injection (AutoDojo) — 直接測量的注入防禦案例:內容篩選器會偵測類似指令的文字(低成本的 AutoDojo 攻擊者可以繞過的啟發式方法,因此 0%-static 篩選器會洩漏 28%),而確定性動作閘門(Progent、DRIFT)移除執行偏離軌跡寫入的能力,並在五個模型中維持有效——這是測試預測的不對稱性,而且在未充分指定的 action-open 工作上會擴大
- Capability Gating Is Not Authorization — fail-closed 的PDP/PEP會移除授權偏離政策呼叫的能力(預設拒絕、錯誤即拒絕),而非限制它;「對格式錯誤輸入 fail open 的政策引擎,會用額外步驟重現該漏洞」正是將這項測試表述為實作規則
待解決的問題#
- 深度防禦傳統上會在「足夠多的摩擦控制措施可以累積成一道屏障」的理論上,堆疊摩擦控制措施。這項測試是否使分層摩擦失效,還是只會將其降級到能力移除之下?
- 有些控制措施對人類而言是摩擦,對代理而言卻是屏障(反之亦然)。這項測試是否以代理為相對對象?對混合人類/代理威脅模型,應如何評估?
資料來源#
- Zero Trust for AI Agents — 「A design test: impossible, not tedious」(Part: principles);在 Phases 3 與 5 以及結尾章節中再次出現
Cited by 27
- Does 'Impossible, Not Tedious' Kill Defense-in-Depth? Layered Friction, Agent-Relativity, and the Frequency Paradox×4
A rate (throttle) is friction. N-per-minute, resettable, delay-based: it narrows the window without…
- Foundation → Enterprise → Advanced: Is the Agent Access-Control Jump a Cliff?×3
Both readings are defensible and the gap is real: basic ABAC (a few attributes — data sensitivity,…
- Agentic Prompt Injection×3
Encoding-based filters and pattern blocklists are friction controls: a patient attacker re-encodes…
- Blast Radius (Agentic)×3
Blast-radius assessment must be run through the Impossible Not Tedious Test: "If your containment…
- Zero Trust for AI Agents×3
Define agent boundaries — unique identity, approved/prohibited actions, escalation triggers, scope…
- Agent Identity and Authentication×2
Foundation — short-lived, narrowly-scoped tokens from an identity provider (OAuth 2.0), expiry in…
- AI-Accelerated Offense×2
Impossible Not Tedious Test — near-zero per-attempt cost is precisely what breaks friction controls
- Autonomous Intrusion×2
Impossible Not Tedious Test (hub) — "many thousands of individual actions" is the observed instance…
- Capability-Gated Model Fallback×2
Anthropic concedes it is "likely impossible to completely prevent universal jailbreaks"; the goal…
- Capability Gating Is Not Authorization×2
Impossible Not Tedious Test — a deterministic fail-closed gate removes the capability to authorize…
- Least Agency×2
Least agency adds a frequency dimension ("how often"), but the framework also says rate limits are…
- LLM-Driven Vulnerability Research×2
Impossible Not Tedious Test — the "friction-based defenses degrade" observation here is turned into…
- MCP Tool Poisoning×2
Impossible Not Tedious Test — the human-consent control that works here (fires at the irreversible…
- Task-Specification Effects in Prompt Injection (AutoDojo)×2
The paper's takeaway: "Real robustness comes from binding the agent's actions to the user's request…
- Agent Context Files
The disclosure is not the vulnerability. The exposure is that an agent reads issue text with…
- Agent Data Injection (ADI)
Impossible Not Tedious Test — user-confirmation dialogs are a friction control that fails (the…
- Classifier Gates vs OS Sandboxing: The Defense-in-Depth Story for Auto Mode and Cowork
The two controls sit on opposite sides of the Impossible Not Tedious Test. The auto-mode classifier…
- Can Models Learn to Separate Instructions from Data? Durable Property vs Training Gap
The deepest reason capability alone can't close it: content-level separation is a target an…
- Memory and Context Poisoning
Impossible Not Tedious Test — the "Opus flags but does not delete" behavior is the test failing in…
- Agent Security
Impossible Not Tedious Test (hub) — Zero Trust design test for agentic security: does a control…
- Non-Malleable Memory Authority (TMA-NM)
Impossible Not Tedious Test — TMA-NM removes the capability for untrusted memory to authorize a…
- Off-Host, Identity-Bound Authorization
Impossible Not Tedious Test — a capability-removing (not friction) control: a deterministic deny at…
- Open Questions Backlog
Impossible Not Tedious Test: Some controls are friction for humans but barriers for agents (or vice…
- Out-of-Band Prompt-Injection Defense
Impossible Not Tedious Test — a deterministic reference monitor removes the capability to authorize…
- Self-Propagating Prompt Injection (AI Worms)
Impossible Not Tedious Test — the customer-side mitigations are pure friction, and the source's own…
- Unsanctioned Action in Capability Evaluations
Impossible Not Tedious Test (hub) — friction priced at zero, observed: audio-CAPTCHA defeat by…
- Write-Then-Trusted
Impossible Not Tedious Test — a denylist that is always one entry short is the archetype of a…
Related articles
- Agentic Prompt Injection
Direct and indirect injection of malicious instructions into an agent; LLMs cannot reliably distinguish information fro…
- Zero Trust for AI Agents
Anthropic's security framework for deploying autonomous agents: trust nothing / verify everything / assume breach, appl…
- Least Agency
OWASP term extending least privilege to agents: constrain not just what an agent can access but what each tool can do,…
- Out-of-Band Prompt-Injection Defense
Second-generation prompt-injection defense enforced outside the model: a deterministic reference monitor mediates tool…
- Capability Gating Is Not Authorization
Agent frameworks ship capability gating (which tools are exposed, schema validity) but no fail-closed per-call authoriz…
