AI Agents Run at Machine Speed. Identity Governance Wasn't Built for That.
Enterprises are deploying autonomous AI agents faster than their identity and access controls can keep up, according to a new SailPoint study highlighted by The Hacker News on 9 October. The Horizons of Identity Security report puts a name — the "AI velocity paradox" — to a mismatch that security architects have been warning about since agent-based workflows moved out of labs and into production.
A note on sourcing: the underlying research is sponsored by SailPoint, an identity governance vendor with a commercial interest in the problem it describes. That does not make the diagnosis wrong, but readers — particularly those weighing procurement decisions — should treat the report's framing as vendor-perspective analysis rather than independent measurement. The Hacker News summary discloses no third-party verification of the study's figures.
The core tension
The report's central finding is a speed mismatch. Organisations are funding AI operations designed to run continuously and autonomously, while the controls meant to keep them accountable still assume a human submits requests at human pace, subject to human review cycles.
In plain language: your AI agents act in milliseconds, but the process that decides whether they're allowed to act still moves at the speed of an approval queue.
Legacy identity and access management (IAM) tooling was built to answer a deceptively simple question — is this employee allowed to do this? — against an assumed world of stable roles, predictable entitlements, and onboarding measured in weeks. Autonomous agents break every one of those assumptions. They spawn, chain tools together, hold credentials that outlive the tasks they were issued for, and operate at volumes no review board could plausibly read through.
Crucially, this is a structural problem, not a cultural one. Time-boxed grants and manager-in-the-loop workflows are design properties of governance platforms, not bad habits of security teams. Bolting agents into that model typically produces either paralysis or, more commonly, broad standing privileges that quietly negate the governance the organisation just bought.
Why it matters here
The report does not address any specific jurisdiction, and readers should be wary of vendors back-attributing local conclusions to global research. But the tension is not abstract for organisations in Hong Kong: it lands against an established regulatory backdrop that already treats unauthorised access and data handling as accountable, auditable failures — including the Personal Data (Privacy) Ordinance for data controllers and supervisory technology-risk expectations from the HKMA for licensed institutions. Boards pushing AI agent pilots into production are, in effect, being asked to sign off on accountability structures that were not designed for non-human actors.
A short checklist for IT managers
If you own identity infrastructure, three questions are worth asking before the next agent deployment review:
- Can you name your agents? Is there an inventory of which autonomous agents exist, which distinct identities they use, and which downstream systems they can reach? Shared service accounts inherited from the systems they were scripted against will collapse your audit trail the moment something goes wrong.
- Are entitlements task-scoped? Grants should be tied to a specific task and duration, not standing permissions that persist between runs. This is the largest departure from legacy IAM design.
- Can you revoke at machine speed? Review cycles measured in weeks cannot govern entities acting in milliseconds. Detection and revocation loops have to compress to match the operating tempo of what they govern.
None of these ideas are novel individually. Their value here is as a single economic argument: organisations are currently collecting the productivity gains of AI autonomy while deferring the governance cost — and that deferred cost compounds. Either security architecture accelerates to meet AI operating speed, or the blast radius of the next identity-driven incident expands to match it.
AI Agent 以機器速度運行，Identity Governance 並非為此而設
企業部署自主 AI agent 的速度，已超出其身份及存取控制（identity and access control）能追上的範圍——這是 SailPoint 一項新研究的結論，該研究於 10 月 9 日獲 The Hacker News 報道。這份名為《Horizons of Identity Security》的報告，為安全架構師一直警告的落差命名——「AI velocity paradox（AI 速度悖論）」——該落差自 agent-based workflow 由實驗室走向 production 環境後便已存在。
關於資料來源：這項研究由 SailPoint 贊助，而該公司正是問題所涉範疇的 identity governance 供應商，對此問題有商業利益。這不代表其診斷有誤，但讀者——尤其是正在考慮採購決定的人士——應將報告的論述框架視為供應商角度的分析，而非獨立測量。The Hacker News 的摘要並未披露任何對該研究數據的第三方驗證。
核心矛盾
報告的核心發現是速度上的錯配。機構正在投入資源建立設計為持續、自主運行的 AI 營運系統，但用以確保其問責性的控制機制，仍然假設由人類以人手節奏提交申請，並經過人手審批流程。
換句話說：你的 AI agent 以毫秒計的速度行事，但決定它們是否獲准行事的流程，仍然以審批 queue 的速度運作。
傳統的身份及存取管理（Identity and Access Management, IAM）工具，是為了回答一個看似簡單的問題而設計——這名員工獲准做這件事嗎？——其預設前提是角色穩定、entitlement 可以預測、onboarding 以數週計。自主 agent 把這些前提逐一打破。它們會自行衍生、串連不同工具、持有超出原有任務範圍仍然有效的憑證，並以任何審核委員會都無法逐一過目的規模運作。
關鍵在於，這是結構性問題，而非文化問題。限時授權（time-boxed grants）及 manager-in-the-loop workflow，是 governance 平台的設計特徵，並非安全團隊的壞習慣。硬將 agent 塞入這套模式，通常只會產生兩種結果：癱瘓不動，或者更常見的情況——授予範圍廣泛的常設權限（standing privileges），無聲地抵銷了機構剛剛採購的 governance。
為何與香港息息相關
報告並沒有針對任何特定司法管轄區，讀者應警惕供應商將全球研究「對號入座」推演出本地結論。但對香港的機構而言，這矛盾並非抽象概念：它落在一個已確立的監管背景之上——該背景早已將未經授權存取及數據處理視為須負責、可審計的違規事項，包括適用於 data controller 的《個人資料（私隱）條例》，以及金管局（HKMA）對持牌機構在 technology risk 監督方面的期望。董事會推動 AI agent 試點進入 production，實際上等同於被要求批准一套並非為「非人」行為者而設計的問責架構。
給 IT 管理員的簡短清單
如果你負責 identity 基礎設施，在下一次 agent 部署審查之前，有三個問題值得追問：
- 你能說出你有哪些 agent 嗎？是否有一份 inventory，列明現有哪些自主 agent、各自使用哪些不同的身份、以及可以觸及哪些下游系統？從原有系統沿用下來的共用 service account，一旦出事便會令你的 audit trail 全面崩塌。
- Entitlement 是否限定於特定任務？授權應綁定於具體任務及時段，而非在多次執行之間持續存在的常設權限。這是與傳統 IAM 設計最大的差別。
- 你能以機器速度撤銷權限嗎？以數週計的審批周期，無法管治以毫秒行事的實體。Detection and revocation loop 必須壓縮，以配合被管治對象的運行節奏。
這些概念單獨來看都並非新鮮事物。它們在此的價值，在於構成一個統一的經濟論證：機構目前正收穫 AI 自主性帶來的生產力，卻同時把 governance 的成本推遲——而這筆推遲的成本會不斷複利累積。安全架構要麼加速跟上 AI 的運行速度，要麼下一宗由身份引發的事故，其影響範圍（blast radius）便會相應擴大。