|
Welcome back! Some identities need to remain for years. Their permissions rarely do. That distinction played out when CISA had to follow one exposed administrator across every environment they could reach. Revoking the credentials found in a public repository was only the first step; the agency also had to discover and rotate access elsewhere. CoSAI approaches the problem from the design side. Its guidance gives an agent a stable identity for attribution, while making each task's credential short-lived and narrowly scoped. The record persists; the permission ends. Transport for London's recovery shows the institutional burden when access cannot be unwound neatly. After a four-day intrusion, 27,000 employees had to reset passwords in person before disrupted digital services could recover. These are not three versions of an AI incident. CISA and TfL involved conventional identity and access failures. They matter because they expose the operating conditions that autonomous systems will inherit: incomplete reach maps, persistent entitlement and revocation designed under pressure. My perspective: accountability persists; authority expires. An organisation proves control when it can establish who sponsored an identity, what it may do, where it can act, what it did and when its permission ended. This week's Big Read, The containment debt of permanent privilege, examines the operating liability created when any part of that chain remains unknown. |
Context update Once access is granted, can the organisation still trace its reach, limit its scope and withdraw it without reconstructing the system under pressure? |
|
Reach defines the response
The obvious response to credentials appearing in a public repository is to revoke what is visible. CISA's May response shows why visibility and reach are different problems. After a contractor's AWS GovCloud credentials appeared on a personal GitHub account, CISA disabled the repository and development environment, removed the contractor's access and replaced the disclosed credentials. The agency then rotated credentials across every environment where the same individual held administrator rights. The leak identified a starting point. The administrator's cross-system reach determined the containment boundary.
|
Identity should persist; authority should not Rather than discovering the authorisation reach during an incident, it really should be part of the access design process. CoSAI's 2026 agentic identity guidance recommends treating an agent as a first-class identity, distinct from the human or role it represents. That identity is stable because attribution needs a durable reference point, but not a permanent credential. For action, CoSAI recommends short-lived entitlement tied to a defined task. Scope should narrow at each downstream hop, and the final system should enforce what the token actually permits instead of relying on an earlier gateway decision. The distinction is important because stable identities preserve the record of who or what acted. Expiring authority prevents that identity from carrying yesterday's permission into tomorrow's task.
|
Recovery tests the shape of authority Transport for London's 2024 breach shows what recovery looks like when compromised internal access cannot be withdrawn selectively. Court records, law-enforcement material and TfL disclosures describe a four-day intrusion followed by an institution-wide reset exercise. All 27,000 employees had to reset their passwords in person. Digital booking and accessibility services were disrupted for weeks, while buses and the Underground continued to operate. Password revocation alone did not preserve transport operations here; what carries over is the scale of authority that had to be unwound before digital recovery could proceed.
|
In Focus The authority chain must exist before the incident The mechanism beneath these cases is not merely credential hygiene. It is the absence of an organisation-controlled authority chain: a live record connecting an identity to its sponsor, purpose, reach, actions, expiry and revocation path. Start with sponsorship. Every agent, integration, service account and administrator should have a named owner who remains responsible for the identity's validity. The record should also identify the person or organisational role on whose behalf it acts. Purpose comes next. A role label such as “automation” or “administrator” does not explain why authority exists. A useful purpose statement names the task, the permitted actions and the condition that ends the permission. Reach must be positive and current. The organisation should know every system, environment and data boundary the identity can access without asking the identity, its holder or a vendor to reconstruct the answer. Action evidence belongs at the systems where work occurs. Gateway logs can show that a request entered. They cannot prove that a downstream resource honoured its scope or record what the final system allowed. Duration is where a temporary grant becomes standing privilege. An expiry date is useful, but task completion is the stronger boundary. Revocation then needs one accountable owner and a tested route across every place the entitlement was accepted. These elements turn accountability from a policy claim into an operating capability. CoSAI's recommendations supply a useful architecture for agents, including narrower downstream scope, final-system enforcement and tamper-evident provenance. Harvard Business Review's analysis of outsourced AI risk adds qualified pressure. Its liability framing varies by jurisdiction, contract and facts; it is not universal law. The practical governance question nevertheless survives: can the deploying organisation reconstruct what acted in its name from evidence it controls? If the answer depends entirely on a supplier's records, accountability is conditional on the supplier's visibility. If expiry cannot be established, the authority chain has no ending. |
This week's Big Read The containment debt of permanent privilege Why effective accountability depends on making authority temporary.
|
Poll of the week Which authority fact would take longest to establish in your organisation today?
|
Cheat Sheet Four artefacts make the authority chain testable before recovery forces the exercise:
|
Standing privilege is deferred containment work. Its daily convenience is visible; its accumulated cost appears only when an organisation must locate, narrow and revoke access faster than its records allow. The size of that bill was fixed before the breach. Each undocumented system, inherited entitlement and missing expiry condition added another dependency to an exercise nobody had scheduled. Good governance keeps identity attributable while making authority disposable. The institution should retain evidence of action, not permission to repeat it indefinitely. |
Choose one agent, service account or administrator this week. Trace every system it can reach, then find the record that says when its authority ends. If the second step takes longer, reply and tell me where the trail stopped. Use the referral page to pass the test to a colleague responsible for access or agent governance. Until next week, David |
