>
volgeMicrosoft · 2027
ZERO TRUST SECURITY · VOICE OF VOLGE

An identity for every agent

Zero Trust for autonomous agents through Microsoft Entra Agent ID

Voice of Volge is a series of books, and the firm's thought leadership on the solutions and the domains Volge specialises in. Each edition takes one of them and sets out what the technology makes possible, what it makes mandatory, and what an operator has to put in place before either is true of them.

Zero Trust for autonomous agents through Microsoft Entra Agent ID

Figures carry their basis and their date. Where published estimates disagree, the range is given and the basis named rather than averaged away, and where a widely quoted number does not survive checking it is corrected here rather than repeated. Every chart plots a figure stated in the text, and nothing is projected beyond what a named source projects.

Contents

  1. 02Executive summary
  2. 03, Anatomy of a rogue agent
  3. 05, Making the invisible workforce countable
  4. 08, Closing the four gaps
  5. 12, Sequencing the controls
  6. 15, Identity in the Gulf and India
  7. 18, The identity leadership agenda
  8. 20Appendices
Voice of Volge is a series of books, and the firm’s thought leadership on the solutions and the domains Volge specialises in.
Microsoft · 2026
Zero Trust SecurityVoice of Volge01 •

Executive summary

A second work­force now reaches into cor­por­ate data, and al­most none of it ap­pears on a payroll or in a joiner-mover-leaver pro­cess or in an iden­tity team's in­vent­ory. Every autonom­ous agent shipped dur­ing 2025 and 2026 holds a cre­den­tial and calls tools. It quer­ies data stores and writes against live sys­tems. Each is a non-human iden­tity (NHI). Most are un­scoped and un­logged, and im­possible to re­voke cleanly, be­cause they were stood up without a first-class iden­tity in the first place.

The title of this paper is an op­er­a­tional ques­tion, not a rhet­or­ical one. In a large en­ter­prise the pop­u­la­tion touch­ing cor­por­ate data is already pre­dom­in­antly non-human, and the human se­cur­ity team can at­trib­ute al­most none of it. Ma­chine iden­tit­ies, agent iden­tit­ies in­cluded, out­num­ber human users by 109 to one across the en­ter­prises sur­veyed, up from 82 to one a year earlier (Palo Alto Net­works, 2026 Iden­tity Se­cur­ity Land­scape, sur­vey of 2,930 cy­ber­se­cur­ity lead­ers). Agent de­ploy­ment is widen­ing that ratio, not nar­row­ing it.

The risk does not sit in the agents. It sits in the asym­metry between how fast they were de­ployed and how slowly the con­trols fol­lowed. Sixty-five per cent of en­ter­prises re­port that an agent has already acted out­side its in­ten­ded scope or au­thor­ised ac­cess, and 29 per cent of those re­port a meas­ur­able ef­fect on the or­gan­isa­tion (En­ter­prise Man­age­ment As­so­ci­ates, Septem­ber 2026). The agent that causes the loss is rarely com­prom­ised and never ma­li­cious. It is over-per­mis­sioned, un­tracked, and trus­ted by de­fault, which is the same fail­ure mode as a stand­ing ser­vice ac­count, scaled and made autonom­ous.

The cent­ral claim of this paper is the fol­low­ing. An agent without a first-class, cryp­to­graphic iden­tity can­not be in­vent­or­ied or scoped, and it can­not be audited or re­voked. It is there­fore not a pro­ductiv­ity asset but an un­metered path into cor­por­ate data wear­ing the badge of one. Se­cur­ing the agent es­tate is an iden­tity pro­gramme be­fore it is any­thing else. In the Mi­crosoft es­tate, the iden­tity is sup­plied by Mi­crosoft Entra Agent ID and op­er­ated through Mi­crosoft Agent 365.

The board-level case in the ref­er­ence flag­ship of this series, From co­pi­lot to auto­pi­lot, settled that iden­tity is the front­line of agent con­trol. This paper is writ­ten for the people who now have to build that front­line, from the iden­tity ad­min­is­trator to the se­cur­ity en­gin­eer and the chief in­form­a­tion se­cur­ity of­ficer's team. It is about the mech­an­ics and how each is con­figured: the agent iden­tity ob­ject, the re­gistry, con­di­tional ac­cess for agents, tool-level least priv­ilege, the cre­den­tial li­fe­cycle, the audit trail and the kill-switch.

The ar­gu­ment runs in five moves, de­veloped across the chapters that fol­low.

A single rogue agent is a chain of five iden­tity omis­sions, and any one of them breaks the chain. The breach does not begin with a clever ad­versary. It be­gins with broad ac­cess gran­ted be­cause scop­ing is harder than grant­ing, a cre­den­tial that never ex­pires, an agent ab­sent from any in­vent­ory, ac­tions logged in no re­view­able form, and ac­cess that is never with­drawn when the task ends. Each omis­sion is an iden­tity con­trol that was skipped. The agent-se­cur­ity prob­lem is the sum of those skipped con­trols.

The fix is to make every agent count­able, be­cause what can­not be enu­mer­ated can­not be gov­erned. An agent with a first-class iden­tity can be placed in a re­gistry and scoped to least priv­ilege. It can be eval­u­ated by policy at each re­quest, then logged and re­voked. An agent without one can be none of these things. Mi­crosoft Entra Agent ID reached gen­eral avail­ab­il­ity in April 2026 and as­signs each agent a gov­erned iden­tity with spe­cial­ised OAuth flows, li­fe­cycle man­age­ment and a full audit trail (Mi­crosoft, 2026).

Zero Trust reaches agents as four con­figured set­tings, not as a slo­gan. Con­di­tional Ac­cess for agents eval­u­ates every token an agent iden­tity ac­quires, scopes policy by ob­ject iden­ti­fier, cus­tom se­cur­ity at­trib­ute, or blue­print, and de­term­ines whether to allow, block or limit the ac­cess or to re­quire fur­ther veri­fic­a­tion (Mi­crosoft Learn, 2026). A policy can block an agent that Iden­tity Pro­tec­tion rates high-risk, a con­di­tion that re­mains in pre­view as at Novem­ber 2026. Least priv­ilege binds to the tool call, not to the agent. The cre­den­tial is short-lived and pur­pose-bound, not a stand­ing secret. These are knobs an iden­tity team turns, and this paper spe­cifies them.

The audit trail and the kill-switch are the floor, and most es­tates are below it. Every au­then­tic­a­tion and token ac­quis­i­tion has to be logged, as does every tool in­voc­a­tion and policy de­cision. The log has to an­swer four ques­tions: what was done, on whose be­half, under what policy and why. Eighty-four per cent of or­gan­isa­tions doubt they could pass a com­pli­ance audit fo­cused on agent be­ha­viour or ac­cess con­trols (Cloud Se­cur­ity Al­li­ance, Feb­ru­ary 2026). The abil­ity to cut an agent's au­thor­ity in­stantly and es­tate-wide is the op­er­a­tional twin of that audit trail, and it is worth­less until it has been tested against a live agent.

The stand­ards landed within eight weeks of each other and they agree. The OWASP Top 10 for Agen­tic Ap­plic­a­tions, pub­lished on 9 Decem­ber 2025, and the Cloud Se­cur­ity Al­li­ance Agen­tic Trust Frame­work, pub­lished on 2 Feb­ru­ary 2026, both put iden­tity first and both co­dify the same pos­ture: no agent trus­ted by de­fault, autonomy earned through demon­strated be­ha­viour, trust con­tinu­ously veri­fied (OWASP, 2025; Cloud Se­cur­ity Al­li­ance, 2026). The Mi­crosoft con­trol set maps onto both, con­trol for con­trol.

Zero Trust SecurityVoice of Volge02 •

A re­gional reader should know the hole in the middle of every res­id­ency ar­gu­ment about agents, and the hole is iden­tity it­self. Mi­crosoft names Mi­crosoft Entra ID as a ser­vice that might store data glob­ally, and Entra mul­ti­factor au­then­tic­a­tion data as data that may be stored glob­ally, as ex­press ex­cep­tions to a ten­ant's chosen geo­graphy, in both the Co­pi­lot Stu­dio and the Dy­nam­ics 365 res­id­ency doc­u­ment­a­tion (Mi­crosoft Learn, pages up­dated 30 April and 7 Oc­to­ber 2026). The iden­tity layer is the one layer Mi­crosoft does not place in re­gion any­where. Agent 365, which car­ries the re­gistry and the agent traces, has no re­gional dial at all and stores and pro­cesses at the ten­ant's home geo­graphy. For a Gulf or In­dian en­ter­prise, the iden­tity and gov­ernance evid­ence for the agent land­scape is there­fore its least loc­al­ised part.

The reg­u­lator, mean­while, re­quires noth­ing spe­cific of it. No in­stru­ment in Saudi Ar­a­bia, the UAE, Qatar, Kuwait, Oman, Bahrain or India reg­u­lates non-human iden­tity as a cat­egory, and none defines a ser­vice ac­count, a work­load iden­tity or a ma­chine cre­den­tial. That is a veri­fied neg­at­ive from ex­haust­ive text search of the primary in­stru­ments, not a gap in the read­ing. The nearest ob­lig­a­tions are an In­dian pay­ments rule on per-ap­plic­a­tion ac­cess con­trol lists and a Saudi bank­ing con­trol on non-per­sonal priv­ileged ac­counts. Chapter five sets out what an op­er­ator does in the ab­sence of a rule, and why the ab­sence is tem­por­ary.

The Mi­crosoft eco­sys­tem is this paper's primary lens be­cause it is the de­fault iden­tity es­tate of the large en­ter­prise, not be­cause the prob­lem is Mi­crosoft-spe­cific. Entra Agent ID brings third-party agents built on plat­forms such as Amazon Web Ser­vices Bed­rock and n8n under the same iden­tity, so the con­trol plane is not con­fined to Mi­crosoft-built agents (Mi­crosoft Learn, 2026). Volge's po­s­i­tion is plain: the iden­tity de­cision is the pre­con­di­tion for every other agent-se­cur­ity con­trol, and it is the de­cision most en­ter­prises have made by de­fault rather than by design. This paper sets out how to make it de­lib­er­ately.

, Anatomy of a rogue agent

how a single identity gap becomes a data breach

The takeaway

A rogue agent is not an exotic event. It is the pre­dict­able out­put of a five-step pipeline in which every step is an iden­tity con­trol that was skipped, and the rate at which it pro­duces breaches is already rising with ad­op­tion. Trace the pipeline and the rem­edy is ob­vi­ous: break it at any step and the breach does not occur. The first task of agent se­cur­ity is to make this pipeline vis­ible in the es­tate, which is an iden­tity task be­fore it is a net­work or data task.

The pipeline, step by step

The mech­an­ism by which agents pro­duce rogue data ex­pos­ure is well evid­enced and struc­tural rather than ac­ci­dental. It runs the same way each time.

An agent is shipped quickly to cap­ture value. To make it use­ful, it is handed broad ac­cess, be­cause scop­ing ac­cess nar­rowly costs more en­gin­eer­ing than grant­ing it widely. The agent is re­gistered in no cent­ral in­vent­ory, be­cause no in­vent­ory for non-human iden­tit­ies ex­ists. Its cre­den­tial is a stand­ing secret with no ex­piry. Its ac­tions land in no log a se­cur­ity ana­lyst would ever read. When the task that jus­ti­fied the agent ends, neither the cre­den­tial nor the ac­cess is with­drawn.

The out­put is an over-per­mis­sioned, un­tracked, im­mor­tal iden­tity with a clear path to cor­por­ate data. Sixty-five per cent of en­ter­prises re­port that an agent has acted out­side its in­ten­ded scope or au­thor­ised ac­cess, which is the ex­pec­ted res­ult when task scope, tool ac­cess, and cre­den­tial life­time are never aligned (En­ter­prise Man­age­ment As­so­ci­ates, Septem­ber 2026). More than half re­port scope vi­ol­a­tions oc­cur­ring oc­ca­sion­ally or more often (Cloud Se­cur­ity Al­li­ance, April 2026, study com­mis­sioned by Zen­ity).

Read the pipeline as a list of con­trols and the dia­gnosis writes it­self. The agent was not re­gistered and not scoped. It was not cre­den­tial-man­aged, and it was neither logged nor re­voked. Each of those five is an iden­tity con­trol. The chapters that fol­low re­store each in turn, and chapter three treats the four that an iden­tity team con­fig­ures dir­ectly.

Why this pipeline is worse for agents than for prior machine identities

Iden­tity teams already man­age non-human iden­tit­ies. Ser­vice ac­counts get ro­tated, cer­ti­fic­ates get man­aged and keys get vaul­ted. Those tech­niques are ne­ces­sary for agents and in­suf­fi­cient, and the reas­ons are spe­cific. Four prop­er­ties of agents widen each step of the pipeline above.

An agent is non-de­term­in­istic, so static scop­ing under-counts its reach. A ser­vice ac­count runs the same code path every time. An agent's be­ha­viour is a func­tion of a model's out­put, which moves with the input. The same agent, given the same nom­inal task, may pick a dif­fer­ent tool or fol­low a dif­fer­ent path on a second run. What an agent can do is there­fore harder to bound by in­spec­tion than con­ven­tional soft­ware, and the broad grant in step two of the pipeline is rarely un­der­stood as broad at the time it is made.

An agent's au­thor­ity is com­pos­ite, so the ef­fect­ive per­mis­sion set is easy to mis­judge. An agent often acts on be­half of a user, in­her­it­ing part of that user's ac­cess, while also hold­ing its

Zero Trust SecurityVoice of Volge03 •

own iden­tity and its own per­mis­sions. The reach is the union of the two, and the union is routinely un­der­es­tim­ated. Mi­crosoft's model sep­ar­ates the cases ex­pli­citly, dis­tin­guish­ing del­eg­ated-ac­cess agents that act on be­half of a user from own-ac­cess agents that op­er­ate with their own iden­tity not tied to a user ses­sion (Mi­crosoft Learn, 2026).

An agent del­eg­ates, so a nar­row grant be­comes broad reach through in­her­it­ance. An agent calls other agents, and those call fur­ther agents. Per­mis­sion in­her­it­ance down a del­eg­a­tion chain is how a tightly scoped task ac­quires a wide blast ra­dius. Ex­cess­ive agency, which the OWASP Top 10 for Agen­tic Ap­plic­a­tions (Decem­ber 2025) names as a primary risk, is most often the product of an under-con­trolled del­eg­a­tion chain rather than a single over-per­mis­sioned agent (OWASP, 2025).

An agent per­sists, so step five of the pipeline is the de­fault rather than the ex­cep­tion. Once de­ployed, an agent and its cre­den­tial typ­ic­ally out­live the task that jus­ti­fied them. Where an agent au­then­tic­ates with a shared key, a human cre­den­tial, or a gen­eric ser­vice ac­count rather than with an iden­tity of its own, the cre­den­tial and en­ti­tle­ment sprawl per­sists in­def­in­itely, and 44 per cent of or­gan­isa­tions re­port re­ly­ing on static API keys to au­then­tic­ate their agents (Cloud Se­cur­ity Al­li­ance, Feb­ru­ary 2026).

What each omission looks like in a live estate

Each of the five omis­sions has an ob­serv­able sig­na­ture, and an iden­tity team can look for all five this week without new tool­ing. The sig­na­tures mat­ter more than the tax­onomy, be­cause they are what a re­view ac­tu­ally finds.

  • No iden­tity. Sign-in re­cords for a work­load re­solve to a human user's ac­count, or to one ser­vice prin­cipal shared by sev­eral agents. The tell is a single cre­den­tial whose activ­ity pat­tern can­not be­long to one actor.
  • No re­gis­tra­tion. The build teams' own lists and the ten­ant query dis­agree. The tell is an agent with re­cent token activ­ity that no in­vent­ory names and no per­son claims.
  • No scop­ing. The agent's grant is a role rather than a list of tool per­mis­sions. The tell is an agent whose doc­u­mented pur­pose is nar­row and whose con­sen­ted scopes are broad.
  • No log­ging. An ac­tion can be seen but not at­trib­uted. The tell is an in­vest­ig­a­tion that stalls at the ques­tion of which agent acted, or on whose be­half.
  • No re­voc­a­tion. Cre­den­tials out­live the task. The tell is an agent hold­ing a cre­den­tial older than the pro­ject that cre­ated it, with no dated test of its dis­able­ment.

Four of the five are meas­ur­able as ra­tios and be­long in a monthly re­port. They are the share of agents hold­ing a ded­ic­ated iden­tity, the share present in the re­gistry, the share scoped at the tool, and the share with a dated re­voc­a­tion test. A team able to pro­duce those four num­bers is already ahead of the 21 per cent of or­gan­isa­tions that main­tain a real-time re­gistry of their agents (Cloud Se­cur­ity Al­li­ance, Feb­ru­ary 2026).

What existing machine-identity practice transfers, and what does not

An iden­tity team is not start­ing from noth­ing, and say­ing so mat­ters, be­cause the op­pos­ite fram­ing in­vites a par­al­lel pro­gramme nobody can staff. Three ex­ist­ing dis­cip­lines trans­fer to agents, each with one qual­i­fic­a­tion.

Secrets man­age­ment trans­fers, with the qual­i­fic­a­tion that a vault solves stor­age rather than life­time. An agent cre­den­tial held safely but never ex­pir­ing is still step four of the pipeline.

Priv­ileged ac­cess man­age­ment trans­fers, with the qual­i­fic­a­tion that its unit is a human ses­sion and an agent has none. The ap­proval work­flow and the ses­sion re­cord are re­usable. The as­sump­tion that a per­son sits be­hind the re­quest is not.

Joiner-mover-leaver trans­fers, and it is the most valu­able of the three, be­cause its whole pur­pose is with­draw­ing ac­cess when a role ends. The qual­i­fic­a­tion is that an agent has no leaver event un­less some­body defines one, so the trig­ger has to be built from the task's com­ple­tion or from the owner's de­par­ture.

What does not trans­fer is the in­spec­tion as­sump­tion. A con­ven­tional non-human iden­tity is bounded by read­ing its code. An agent's reach is a func­tion of a model's out­put and of the tools it may call, which is why the con­trol moves to the tool call and to the tem­plate rather than sit­ting on the in­stance.

The population behind the pipeline, quantified

The pipeline scales with the size of the non-human pop­u­la­tion, and that pop­u­la­tion has crossed a threshold that changes what iden­tity man­age­ment is for. Iden­tity is no longer prin­cip­ally about people.

The re­por­ted ra­tios vary by es­tate type, and the vari­ation is it­self the find­ing. Ma­chine iden­tit­ies out­num­ber human users by 109 to one across the 2,930 en­ter­prises sur­veyed for the 2026 Iden­tity Se­cur­ity Land­scape, with 79 agent iden­tit­ies per human in­side that fig­ure (Palo Alto Net­works, 2026). Entro Labs meas­ured 144 to one in cloud-nat­ive and DevOps en­vir­on­ments in the first half of 2025, and ManageEngine re­ports a sec­tor-level max­imum of up to 500 to one, with nearly half of the or­gan­isa­tions it sur­veyed above 100 to one (ManageEngine, 2026 Iden­tity Se­cur­ity Out­look).

Zero Trust SecurityVoice of Volge04 •
Estate typeNon-human to human identity ratioSource basis
Enterprise average109 : 1Palo Alto Networks, 2026
Of which, agent identities79 : 1Palo Alto Networks, 2026
Cloud-native / DevOps144 : 1Entro Labs, first half of 2025
Sector maximumUp to 500 : 1ManageEngine, 2026

Ra­tios are drawn from dif­fer­ing sur­vey bases and are not strictly com­par­able. The con­sist­ent sig­nal is that non-human iden­tit­ies out­num­ber human users by at least one, and fre­quently two, or­ders of mag­nitude.

By the num­bers , the iden­tity in­ver­sion. Across the en­ter­prises sur­veyed for the 2026 Iden­tity Se­cur­ity Land­scape, ma­chine iden­tit­ies out­num­ber human users by 109 to 1, and agent iden­tit­ies alone by 79 to 1 (Palo Alto Net­works, 2026). The human se­cur­ity team can usu­ally ac­count for the human iden­tit­ies. It can ac­count for al­most none of the ma­chine iden­tit­ies, and least of all the agents, which are the fast­est-grow­ing slice and the one whose reach is hard­est to bound.

The proof the pipeline is running: the audit gap

The single clearest meas­ure that this pipeline is live in most es­tates is that or­gan­isa­tions can­not demon­strate con­trol of their agents to an aud­itor. Eighty-four per cent doubt they could pass a com­pli­ance audit fo­cused on agent be­ha­viour or ac­cess con­trols (Cloud Se­cur­ity Al­li­ance, Feb­ru­ary 2026). Only 18 per cent are highly con­fid­ent that their ex­ist­ing iden­tity and ac­cess man­age­ment can man­age agent iden­tit­ies, with a fur­ther 35 per cent mod­er­ately and 29 per cent slightly con­fid­ent, while 21 per cent main­tain a real-time re­gistry of their agents and 28 per cent can trace agent ac­tions to a human across all en­vir­on­ments (Cloud Se­cur­ity Al­li­ance, Feb­ru­ary 2026, sur­vey of 285 IT and se­cur­ity pro­fes­sion­als).

CapabilityShare of organisationsSource
Doubt they could pass an audit of agent behaviour or access84%Cloud Security Alliance, February 2026
Highly confident existing IAM can manage agent identities18%Cloud Security Alliance, February 2026
Maintain a real-time registry of agents21%Cloud Security Alliance, February 2026
Can trace agent actions to a human across all environments28%Cloud Security Alliance, February 2026

The sur­vey be­hind these fig­ures, Se­cur­ing Autonom­ous AI Agents, was com­mis­sioned and fin­anced by Strata Iden­tity and fielded on­line in Septem­ber and Oc­to­ber 2025; the Cloud

Se­cur­ity Al­li­ance ana­lysed it and pub­lished the res­ults on 5 Feb­ru­ary 2026.

These four num­bers de­scribe the same gap from four angles. The en­ter­prise stood up a non-human work­force faster than it built the con­trols to gov­ern it, and the pipeline of sec­tion 2.1 has been run­ning in the in­ter­val. Clos­ing it is the sub­ject of the rest of this paper.

The implication for the security agenda

The agent-se­cur­ity prob­lem re­duces to a non-human-iden­tity prob­lem, and the proof is that every step of the breach pipeline is an iden­tity con­trol that was skipped: no iden­tity, no re­gis­tra­tion, no scop­ing, no log­ging, no re­voc­a­tion. None of these is a net­work con­trol and none is a data-loss con­trol in the first in­stance. The first move is there­fore to give every agent a first-class iden­tity and put it in a re­gistry, which makes the pop­u­la­tion count­able and the pipeline break­able. That is the sub­ject of chapter two.

, Making the invisible workforce countable

identity as the control plane

The takeaway

What can­not be enu­mer­ated can­not be gov­erned, so iden­tity is the con­trol plane and not one con­trol among sev­eral. An agent with a first-class, cryp­to­graphic iden­tity be­comes gov­ern­able in every other di­men­sion; an agent without one is gov­ern­able in none. Mi­crosoft Entra Agent ID sup­plies that iden­tity in the Mi­crosoft es­tate through a spe­cific ob­ject model. That model has an agent iden­tity blue­print and its prin­cipal, the agent iden­tity in­stan­ti­ated from the blue­print, and an op­tional user ac­count for the agent, and an iden­tity team must un­der­stand it be­fore it con­fig­ures any­thing down­stream. The agent re­gistry, con­sol­id­ated into Mi­crosoft Agent 365, is the in­vent­ory that turns the in­vis­ible work­force into a coun­ted one.

What "first-class" buys, and what its absence costs

A first-class iden­tity is one the iden­tity plat­form re­cog­nises nat­ively, as it re­cog­nises a user or a work­load, rather than a secret bolted onto a human ac­count. The dis­tinc­tion is the whole of the mat­ter, be­cause every down­stream con­trol de­pends on it.

Read those four prop­er­ties against the breach pipeline of chapter one and each maps to a step the pipeline skipped: cryp­to­graphic cre­den­tials re­place the stand­ing secret, enu­mer­a­tion ends the miss­ing in­vent­ory, policy-ad­dress­ab­il­ity

Zero Trust SecurityVoice of Volge05 •

en­ables scop­ing, li­fe­cycle man­age­ment forces with­drawal at task end. An agent that au­then­tic­ates with a human user's cre­den­tials, or with a gen­eric ser­vice ac­count shared across many agents, has none of the four. It can­not be told apart from the human or from its sib­lings, can­not be policy-ad­dressed on its own, and can­not be re­voked without break­ing everything that shares the cre­den­tial. That con­fig­ur­a­tion is still com­mon: 44 per cent of or­gan­isa­tions re­port au­then­tic­at­ing their agents with static API keys (Cloud Se­cur­ity Al­li­ance, Feb­ru­ary 2026).

A first-class agent iden­tity has four prop­er­ties. It is cryp­to­graphic, au­then­tic­at­ing through man­aged cre­den­tials and stand­ard pro­to­cols rather than a shared key. It is enu­mer­able, ap­pear­ing in a dir­ect­ory where it can be found and at­trib­uted and coun­ted. It is policy-ad­dress­able, so ac­cess policy can tar­get it spe­cific­ally. It is li­fe­cycle-man­aged, so it can be cre­ated, then gov­erned and de­com­mis­sioned as a man­aged ob­ject.

The Entra Agent ID object model

Mi­crosoft Entra Agent ID reached gen­eral avail­ab­il­ity in April 2026 and ex­tends Mi­crosoft Entra iden­tity and ac­cess man­age­ment to AI agents (Mi­crosoft, 2026). The ob­ject model has to be un­der­stood be­fore any­thing else is con­figured. Con­di­tional ac­cess, gov­ernance and audit­ing all ref­er­ence these ob­jects by name. The model below is drawn from Mi­crosoft's own defin­i­tions (Mi­crosoft Learn, 2026).

ObjectWhat it isFunction
Agent identity blueprintAn object in Microsoft Entra ID serving as the template and authentication foundation for one or more agent identitiesHolds the credentials and acquires tokens on behalf of every agent identity created from it; policy applied to it covers them all, including identities created later
Agent identity blueprint principalThe object recording a blueprint's presence in the tenantLets the blueprint acquire tokens and appear in audit logs; executes only the creation of agent identities and agents' user accounts, and cannot itself access resources
Agent identityThe primary identity an agent uses to authenticateHolds no credentials of its own and authenticates with tokens issued by its blueprint; the actor in agentic tasks
Agent's user accountAn optional account pairing one to one with an agent identityUsed only where the agent must reach systems that require a user object, such as a mailbox or a Teams channel; it does not replace the agent identity, and it can hold neither passwords nor privileged administrative roles

The plat­form au­then­tic­ates and com­mu­nic­ates over stand­ard pro­to­cols , OAuth 2.0 for au­thor­isa­tion, the Model Con­text Pro­tocol (MCP) for tool con­nec­tion, and the Agent-to-Agent (A2A) pro­tocol for inter-agent com­mu­nic­a­tion (Mi­crosoft Learn, 2026). Stand­ard pro­to­cols are what let the same

iden­tity be re­cog­nised across Mi­crosoft and non-Mi­crosoft plat­forms.

The blue­print-to-iden­tity re­la­tion­ship is what lets the model scale. A blue­print is au­thored once, car­ry­ing its se­cur­ity policy, and in­di­vidual agent iden­tit­ies are in­stan­ti­ated from it in a par­ent-child re­la­tion­ship (Mi­crosoft Learn, 2026). A policy that tar­gets a blue­print ap­plies to every agent iden­tity cre­ated from it, in­clud­ing those cre­ated later, which is how an iden­tity team gov­erns thou­sands of agents without touch­ing each one. This is the an­swer to the non-de­term­in­ism prob­lem of chapter one: scope is fixed at the tem­plate rather than chased per in­stance.

One gov­ernance-design point is worth fix­ing in mind. The blue­print prin­cipal can cre­ate agent iden­tit­ies but can­not it­self ac­cess re­sources, and its cre­ation calls are de­lib­er­ately ex­cluded from con­di­tional-ac­cess eval­u­ation (Mi­crosoft Learn, 2026). The con­trol sur­face is there­fore the agent iden­tity and the agent's user ac­count , the ob­jects that ac­quire tokens and reach data. Scope policy to those, not to the blue­print prin­cipal. Mi­crosoft also lay­ers an ad­min­is­trat­ive model over these ob­jects, in which own­ers hold the tech­nical con­fig­ur­a­tion and spon­sors hold the busi­ness ac­count­ab­il­ity, while a man­ager is the human des­ig­nated for an agent's user ac­count (Mi­crosoft Learn, 2026).

Delegated access versus own access: the most consequential choice

The most con­sequen­tial single choice in agent iden­tity is whether an agent acts on be­half of a user or in its own right. The two carry dif­fer­ent per­mis­sion se­mantics, dif­fer­ent blast radii, and dif­fer­ent con­trol ma­tur­ity.

A del­eg­ated-ac­cess agent acts on be­half of a signed-in user and ac­quires ac­cess in that user's con­text, through the on-be­half-of flow. Con­di­tional Ac­cess eval­u­ates the policies as­signed to that user rather than to the agent, so se­lect­ing an agent iden­tity as the policy sub­ject does not cover the re­quests an agent makes on a user's be­half (Mi­crosoft Learn, 2026). The agent in­her­its the user's policy pos­ture , a con­veni­ence and a li­ab­il­ity at once, be­cause the agent then car­ries the user's reach into every tool it touches.

An own-ac­cess agent runs with its own iden­tity, tied to no user ses­sion, and op­er­ates autonom­ously. Mi­crosoft doc­u­ments this as ap­plic­a­tion-only ac­cess, with its own li­cens­ing pre­requis­ites and step-by-step policy con­fig­ur­a­tion, and no pre­view label at­taches to the ac­cess pat­tern (Mi­crosoft Learn, 2026). This is the harder case: the agent's au­thor­ity is en­tirely its own and must be scoped from first prin­ciples rather than in­her­ited, which is pre­cisely where the broad grant of the breach pipeline takes hold.

Zero Trust SecurityVoice of Volge06 •

The design rule fol­lows dir­ectly. Prefer del­eg­ated ac­cess wherever a clear human prin­cipal ex­ists and the agent should not ex­ceed that per­son's reach. Re­serve own-ac­cess iden­tit­ies for genu­inely autonom­ous, event-triggered agents, and then scope them to least priv­ilege ex­pli­citly. An own-ac­cess agent handed broad stand­ing au­thor­ity is the chapter-one pipeline re­con­sti­t­uted in a single ob­ject. One con­sequence of the rule has to be stated. In on-be­half-of flows Mi­crosoft at­trib­utes risky activ­ity to the user rather than to the agent, and the pub­lished agent risk de­tec­tions apply only to autonom­ous agent activ­ity, so pre­fer­ring del­eg­ated ac­cess nar­rows the agent-spe­cific risk sig­nal the se­cur­ity team can act on (Mi­crosoft Learn, 2026).

Classifying a real estate

The del­eg­ated-or-own de­cision is simple in prin­ciple and con­tested in prac­tice, be­cause most agents are built be­fore any­one asks. Ap­plied to the pat­terns that ac­tu­ally ap­pear, the rule re­solves quickly.

  • An as­sist­ant in­side a pro­ductiv­ity ap­plic­a­tion, act­ing when a per­son asks. Del­eg­ated. The human prin­cipal is ob­vi­ous, and the agent should not ex­ceed that per­son's reach.
  • A ser­vice agent hand­ling a cus­tomer's ticket on be­half of a named op­er­ator. Del­eg­ated, with the op­er­ator's own reach de­lib­er­ately nar­rowed, be­cause in­her­it­ance car­ries their full ac­cess into every tool the agent touches.
  • A sched­uled re­con­cili­ation or mon­it­or­ing agent with no human in the trans­ac­tion. Own ac­cess, scoped from first prin­ciples, since there is no prin­cipal to in­herit from.
  • An agent in­voked by an­other agent. De­cide ex­pli­citly, be­cause the de­fault is the one that hurts. Either the chain car­ries the ori­gin­at­ing user's con­text, or the re­ceiv­ing agent acts in its own right under its own nar­rower grant.
  • An agent that must reach a mail­box or a chan­nel. An agent's user ac­count paired to the agent iden­tity, re­mem­ber­ing that the ac­count can hold neither pass­words nor priv­ileged ad­min­is­trat­ive roles, and that policy has to tar­get it sep­ar­ately.

Re­cord the clas­si­fic­a­tion in the re­gistry rather than in the build doc­u­ment­a­tion, be­cause policy scop­ing de­pends on it and the two sub­jects are gov­erned by dif­fer­ent policies. Re-ex­am­ine it whenever an agent's pur­pose changes. That is the mo­ment a del­eg­ated agent quietly ac­quires a sched­ule and be­comes an own-ac­cess agent nobody re-scoped.

The re­gistry shipped first as a stan­dalone Entra blade and was con­sol­id­ated into Mi­crosoft Agent 365 on 1 May 2026, mak­ing Agent 365 the single con­trol plane for agent vis­ib­il­ity, gov­ernance, and se­cur­ity across the ten­ant (Mi­crosoft, 2026).

Through it a se­cur­ity team can ask which agents exist, what iden­tit­ies they use, what data they touch, where they run, which mod­els they de­pend on, and how their ac­tions are audited (Mi­crosoft, 2026). Those six ques­tions are the in­verse of the six things the breach pipeline left un­known. One op­er­a­tional con­sequence be­longs with the stage-one plan of chapter four. The earlier re­gistry Graph API is being de­prec­ated in fa­vour of an Agent 365 API, and agents re­gistered through the earlier API will have to be re­gistered again (Mi­crosoft Learn, 2026).

The registry: turning the population into an inventory

An iden­tity that is not in­vent­or­ied is not gov­erned, be­cause it can­not be found. The agent re­gistry is the in­vent­ory of every agent in the ten­ant, and it is the lit­eral an­swer to the ques­tion that titles this paper.

The re­gistry is only use­ful if it re­cords both in-house and third-party agents, be­cause shadow agents are the largest ex­pos­ure. Entra Agent ID brings third-party agents built on plat­forms such as Amazon Web Ser­vices Bed­rock and n8n under a gov­erned iden­tity through the Mi­crosoft Entra Auth soft­ware de­vel­op­ment kit or work­load iden­tity fed­er­a­tion, re­gard­less of where the agent was built (Mi­crosoft Learn, 2026). An es­tate that re­gisters only Mi­crosoft-built agents leaves its biggest gap un­meas­ured and the pipeline run­ning un­der­neath it.

How the inventory is actually built

A re­gistry is pop­u­lated by re­con­cili­ation, not by in­vit­a­tion. Ask­ing teams to re­gister their agents finds the com­pli­ant ones, and the ex­pos­ure sits in the rest. Work four sources against each other and in­vest­ig­ate every dif­fer­ence.

  • The plat­form's own view. Agents vis­ible in the Agent 365 re­gistry across Co­pi­lot Stu­dio, SharePoint, Agent Builder, Mi­crosoft Foundry and the Mi­crosoft 365 Agents Toolkit, to­gether with the named non-Mi­crosoft plat­forms the re­gistry cov­ers.
  • The dir­ect­ory. Ser­vice prin­cipals and work­load iden­tit­ies with re­cent token activ­ity, read from sign-in logs rather than from a con­fig­ur­a­tion ex­port.
  • The money. Sub­scrip­tions, API keys and model en­d­points vis­ible in billing and in the pro­cure­ment re­cord, which finds agents built out­side the sanc­tioned plat­forms.
  • The net­work. Out­bound calls to model and agent-plat­form en­d­points, which finds the agents ap­pear­ing in none of the first three.
Zero Trust SecurityVoice of Volge07 •

Each re­cord needs five fields be­fore it is use­ful. They are the iden­tity it au­then­tic­ates with, the owner and the spon­sor, whether it holds del­eg­ated or own ac­cess, the tools it may call, and the clas­si­fic­a­tion of the data it touches. A re­cord miss­ing the last field can­not be risk-ranked. A re­cord miss­ing the third can­not be policy-scoped cor­rectly, be­cause a policy tar­get­ing agent iden­tit­ies does not cover an agent's user ac­count.

Two op­er­a­tional traps be­long here rather than in the rol­lout plan. The earlier re­gistry Graph API is being de­prec­ated in fa­vour of an Agent 365 API, and agents re­gistered through the earlier in­ter­face will have to be re­gistered again (Mi­crosoft Learn, 2026). And gov­ernance ac­tions in the re­gistry, ap­prov­ing agent re­quests and as­sign­ing own­er­ship, are re­stric­ted to the AI Ad­min­is­trator and Global Ad­min­is­trator roles (Mi­crosoft Learn, page up­dated 23 Oc­to­ber 2026). Every other role can see a gov­ernance gap and do noth­ing about it, which is a design de­cision worth know­ing be­fore a team dis­trib­utes the work.

What the registry is for, day to day

A re­gistry earns its cost through the ques­tions it an­swers quickly, and six recur in in­cid­ents. Which agents exist, what iden­tit­ies they use, what data they touch, where they run, which mod­els they de­pend on, and how their ac­tions are audited (Mi­crosoft, 2026). Those six are the in­verse of what the breach pipeline left un­known.

In steady state the re­gistry drives four routine tasks. It sup­plies the pop­u­la­tion for the monthly re­port, in­clud­ing the agents found out­side it. It car­ries the ap­proval at­trib­ute that con­di­tional ac­cess reads, so a re­gistry entry is also a policy de­cision. It holds the owner and the spon­sor an in­vest­ig­a­tion has to reach within the hour. And it is the list a joiner-mover-leaver pro­cess reads when an owner leaves, which is the only re­li­able de­fence against the orphaned agent.

Keep it cur­rent by mak­ing it load-bear­ing rather than by audit­ing it. Where the ap­proval at­trib­ute lives in the re­gistry and policy blocks any­thing without it, an un­re­gistered agent simply does not work, and the in­vent­ory main­tains it­self. Where re­gis­tra­tion is a par­al­lel pa­per­work ex­er­cise, it de­cays in­side a quarter.

By the num­bers , the vis­ib­il­ity pre­con­di­tion. Mi­crosoft Entra Agent ID reached gen­eral avail­ab­il­ity in April 2026; Mi­crosoft Agent 365 reached gen­eral avail­ab­il­ity on 1 May 2026; the agent re­gistry con­sol­id­ated into Agent 365 on the same date (Mi­crosoft, 2026). The con­trol plane is a shipped product, not a roadmap item. What is miss­ing in most es­tates is ad­op­tion, not avail­ab­il­ity.

The implication for the security agenda

Iden­tity is the con­trol plane and the re­gistry is its first in­stru­ment. An agent given a first-class Entra Agent ID iden­tity, re­cor­ded in the Agent 365 re­gistry and cor­rectly clas­si­fied as del­eg­ated-ac­cess or own-ac­cess, is an agent that can now be scoped and eval­u­ated, logged and re­voked. The agent that lacks these can­not be gov­erned at all, and stays on the breach pipeline. Scop­ing and eval­u­ation, log­ging and re­voc­a­tion are the four mech­an­ics an iden­tity team op­er­ates against a coun­ted es­tate, and they are the sub­ject of chapter three.

, Closing the four gaps

scope, conditional access, audit, and revocation

The takeaway

A coun­ted es­tate is not yet a gov­erned one. Four con­figured mech­an­ics close the four gaps the breach pipeline opened. Least priv­ilege binds to the tool call and the cre­den­tial, not to the agent. Con­di­tional Ac­cess eval­u­ates every token an agent ac­quires, scopes policy by at­trib­ute or blue­print, and blocks high-risk agents. Audit­ing cap­tures every au­then­tic­a­tion, in­voc­a­tion, and policy de­cision in a re­view­able form. Re­voc­a­tion with­draws an agent's au­thor­ity in­stantly and es­tate-wide. An iden­tity team con­fig­ures all four, and this chapter spe­cifies how.

The posture, and why principle alone changes nothing

The pos­ture now co­di­fied across the in­dustry is that no agent is trus­ted by de­fault, re­gard­less of pur­pose or claimed cap­ab­il­ity, and that trust is earned through demon­strated be­ha­viour and con­tinu­ously veri­fied through mon­it­or­ing (Cloud Se­cur­ity Al­li­ance, 2026). The prin­ciple is cor­rect and inert until it is turned into set­tings. The four mech­an­ics below are those set­tings , con­fig­ur­a­tions, not as­pir­a­tions , and each one closes a named step of the chapter-one pipeline.

The trans­la­tion fol­lows the Cloud Se­cur­ity Al­li­ance Agen­tic Trust Frame­work, pub­lished on 2 Feb­ru­ary 2026. It names five core ele­ments re­quired at every ma­tur­ity level, from iden­tity and be­ha­viour mon­it­or­ing to data gov­ernance, seg­ment­a­tion and in­cid­ent re­sponse. Con­trols scale as the agent earns autonomy (Cloud Se­cur­ity Al­li­ance, 2026). Iden­tity, es­tab­lished in chapter two, is the pre­con­di­tion; the other four as­sume it. The frame­work's ma­tur­ity model runs from In­tern to Prin­cipal across four levels, with agents start­ing read-only and earn­ing autonomy through demon­strated trust­wor­thi­ness across five gates (Cloud Se­cur­ity Al­li­ance, 2026).

Zero Trust SecurityVoice of Volge08 •

Gap one , least privilege at the tool call and the credential

Least priv­ilege for agents does not live at the agent. It lives at each tool the agent calls and at the cre­den­tial the agent holds. This is the cor­rec­tion to the most com­mon agent-se­cur­ity error: scop­ing the agent once at de­ploy­ment and call­ing the task done. That single grant is step two of the breach pipeline.

Scope the tool, not the agent. An agent's ef­fect­ive au­thor­ity is the union of the tools it can in­voke. Ex­cess­ive agency arises when an agent holds more tool per­mis­sions than its task re­quires (OWASP, 2025). The con­trol is to grant each agent only the spe­cific tool per­mis­sions its task needs, and to con­strain per­mis­sion in­her­it­ance down del­eg­a­tion chains so a sub-agent can­not ex­ceed its par­ent's scope. The Cloud Se­cur­ity Al­li­ance frame­work re­quires least priv­ilege at every step of re­curs­ive del­eg­a­tion, not only at the top, which dir­ectly coun­ters the del­eg­a­tion-chain reach de­scribed in chapter one (Cloud Se­cur­ity Al­li­ance, 2026).

Make the cre­den­tial short-lived and pur­pose-bound. The stand­ing secret is the dur­able ex­pos­ure and step four of the pipeline. The con­trol is just-in-time, pur­pose-bound ac­cess: a tem­por­ary, scoped cre­den­tial that ex­pires auto­mat­ic­ally when the task com­pletes. The Cloud Se­cur­ity Al­li­ance frame­work re­quires just-in-time pro­vi­sion­ing and auto­matic cre­den­tial ex­pir­a­tion, yet most or­gan­isa­tions still pro­vi­sion agent cre­den­tials at de­ploy­ment and never re­visit them (Cloud Se­cur­ity Al­li­ance, 2026). Auto-ex­piry also closes step five, with­draw­ing ac­cess at task end without re­ly­ing on any­one to re­mem­ber.

Making just-in-time credentials work

Just-in-time cre­den­tials are the highest-value change on the list and the one most often de­ferred, be­cause a stand­ing secret is usu­ally what keeps a fra­gile in­teg­ra­tion up. Re­pla­cing it is an en­gin­eer­ing task with a pre­dict­able set of obstacles.

Start by clas­si­fy­ing agents by what the cre­den­tial is for. An agent triggered by an event and fin­ish­ing in seconds needs a token is­sued per task and noth­ing more. An agent run­ning a long job needs re­fresh in­side the job rather than a cre­den­tial out­liv­ing it. An agent reached by an­other agent needs its au­thor­ity de­rived from the caller rather than held in its own right. Those pat­terns cover most of a real es­tate, and each has a dif­fer­ent an­swer.

Three obstacles recur. Leg­acy tar­gets that ac­cept only a static key, which is solved by put­ting a broker in front of them rather than by ex­empt­ing the agent. Sched­uled work that out­lives any sens­ible token life­time, which is solved by

short­en­ing the unit of work rather than length­en­ing the cre­den­tial. And own­er­ship, be­cause the cre­den­tial usu­ally be­longs to a team that did not ask for the change and sees only the risk of break­ing some­thing that cur­rently works.

Se­quence the con­ver­sion by ex­pos­ure rather than by ease. Con­vert first the agents hold­ing write ac­cess to a sys­tem of re­cord. The meas­ure of pro­gress is the count of stand­ing secrets re­main­ing, re­por­ted until it reaches zero. Forty-four per cent of or­gan­isa­tions re­port au­then­tic­at­ing their agents with static API keys (Cloud Se­cur­ity Al­li­ance, Feb­ru­ary 2026), so the start­ing count in most es­tates is high, and the trend is the only hon­est meas­ure of the work.

The con­trast between pre­vail­ing and re­quired prac­tice is set out below.

DimensionPrevailing practiceLeast-privilege practice
CredentialStanding secret, no expiryJust-in-time, auto-expiring, purpose-bound
Scope unitThe agent, set once at deploymentEach tool call, evaluated per task
DelegationInherited broadly down the chainConstrained at every step; sub-agent ≤ parent
Permission reviewNever revisitedContinuous; access withdrawn on task completion
Identity basisShared key or human credentialFirst-class agent identity

Sources: OWASP Top 10 for Agen­tic Ap­plic­a­tions, Decem­ber 2025; Cloud Se­cur­ity Al­li­ance Agen­tic Trust Frame­work, Feb­ru­ary 2026.

Gap two , Conditional Access on every token

Con­di­tional Ac­cess is the policy en­gine that eval­u­ates each ac­cess re­quest in real time against con­text and risk. For agents it treats the agent as a first-class iden­tity and eval­u­ates its re­quests as it eval­u­ates a user's, with agent-spe­cific logic (Mi­crosoft Learn, 2026). The mech­an­ics are pre­cise and con­fig­ur­able, and an iden­tity team should know them ex­actly.

What is eval­u­ated. Con­di­tional Ac­cess fires whenever Mi­crosoft Entra ID is­sues or re­freshes an ac­cess token for an agent iden­tity or for an agent's user ac­count (Mi­crosoft Learn, 2026). Some re­sources ad­di­tion­ally sup­port Con­tinu­ous Ac­cess Eval­u­ation, which can trig­ger near-real-time en­force­ment and is what brings the kill-switch of sec­tion 4.5 closer to im­me­di­ate. Eval­u­ation de­lib­er­ately does not fire when a blue­print prin­cipal ac­quires a token only to cre­ate an agent iden­tity, nor on in­ter­me­di­ate token-ex­change calls that reach no re­source. The con­trol sur­face is the token ac­quis­i­tion that touches data , the exact point the breach pipeline left un­gov­erned.

What con­di­tions are as­sessed. The prin­cipal con­di­tion is agent risk, scored high or me­dium or low from Mi­crosoft

Zero Trust SecurityVoice of Volge09 •

Entra ID Pro­tec­tion sig­nals, and in pre­view as at Novem­ber 2026 (Mi­crosoft Learn, 2026). The ac­cess con­trols are to allow, block or limit ac­cess, or to re­quire fur­ther veri­fic­a­tion, and the doc­u­mented agent con­trols in­clude mul­ti­factor au­then­tic­a­tion be­fore a user au­thor­ises an agent, a com­pli­ant device, and a com­pli­ant net­work (Mi­crosoft Learn, 2026). Policies can run re­port-only first to sim­u­late im­pact be­fore en­force­ment, then be toggled on (Mi­crosoft Learn, 2026).

How policy is scoped. Policy can tar­get all agent iden­tit­ies in the ten­ant, spe­cific agent iden­tit­ies by ob­ject iden­ti­fier, agent iden­tit­ies grouped by cus­tom se­cur­ity at­trib­ute, agent iden­tit­ies grouped by their blue­print, or agent users, whether all of them or named ones, and Mi­crosoft has re­placed the flat se­lec­tion list with an en­hanced ob­ject picker (Mi­crosoft Learn, 2026). The cus­tom-se­cur­ity-at­trib­ute method is the re­com­men­ded ap­proach for an ap­proval work­flow. An il­lus­trat­ive at­trib­ute such as AgentApprovalStatus car­ries val­ues such as In_Review, HR_Approved, or IT_Approved, and policy blocks every agent ex­cept those car­ry­ing an ap­proved value (Mi­crosoft Learn, 2026). Two doc­u­mented bound­ar­ies trap ex­actly this design. A policy scoped through the Users as­sign­ment does not apply to agent user ac­counts, and a policy tar­get­ing agent iden­tit­ies does not cover an agent's user ac­count, so an ap­proval gate needs a policy for each sub­ject (Mi­crosoft Learn, 2026).

Two ca­non­ical policies be­long first in any con­fig­ur­a­tion (Mi­crosoft Learn, 2026). The first blocks every agent ex­cept those car­ry­ing an ap­proved cus­tom se­cur­ity at­trib­ute, put­ting an ap­proval gate across the es­tate. The second blocks any agent that Iden­tity Pro­tec­tion rates high-risk, pulling a com­prom­ised or an­om­al­ous agent off re­source ac­cess auto­mat­ic­ally. The second car­ries the heav­ier li­cence, be­cause risk-based Con­di­tional Ac­cess for agents re­quires Entra ID P2 or Mi­crosoft 365 E5 along­side Agent 365, and the agent-risk con­di­tion it de­pends on is still in pre­view. De­ploy both in re­port-only mode, re­view im­pact, then pro­mote to en­force­ment.

Gap three , audit that survives an auditor

An ac­tion logged in no re­view­able form did not, for gov­ernance pur­poses, hap­pen in a gov­ern­able sys­tem. The audit trail is the evid­en­tiary base of the en­tire pos­ture and the pre­cise gap that 84 per cent of or­gan­isa­tions doubt they could close (Cloud Se­cur­ity Al­li­ance, 2026). It is also step five of the breach pipeline , the ac­tion no ana­lyst would ever re­view.

Where it lands. Agent activ­ity sur­faces across the Entra sign-in and audit logs with agent-spe­cific fil­ter­ing. Agent iden­tit­ies act­ing as the actor ap­pear in ser­vice prin­cipal sign-in logs; agent users ap­pear in non-in­ter­act­ive user sign-ins; users

in­vok­ing agents ap­pear in in­ter­act­ive user sign-ins (Mi­crosoft Learn, 2026). Every agent ac­tion is logged with a unique Agent iden­ti­fier, which is what at­trib­utes an ac­tion to a spe­cific agent rather than to a shared ac­count , the at­tri­bu­tion the breach pipeline des­troyed at step three (Mi­crosoft, 2026).

What must be logged. Every agent au­then­tic­a­tion and token ac­quis­i­tion has to be cap­tured, and so does every tool in­voc­a­tion and policy de­cision. The re­cord cov­ers not only what happened but on whose be­half, under what policy and why. All agent au­then­tic­a­tion and activ­ity is logged in Entra for com­pli­ance and audit (Mi­crosoft Learn, 2026).

How it joins the wider es­tate. Agent activ­ity is cap­tured auto­mat­ic­ally for audit and clas­si­fic­a­tion, and the or­gan­isa­tion's ex­ist­ing data-loss-pre­ven­tion policies, sens­it­iv­ity la­bels, and com­pli­ance con­trols in Mi­crosoft Pur­view apply to agents (Mi­crosoft Learn, 2026). When a sus­pi­cious ac­tion leaves the ten­ant via a del­eg­ated agent, the rel­ev­ant Pur­view audit re­cords, the agent's ser­vice-prin­cipal sign-in logs, and the De­fender in­spec­tion ver­dict as­semble into a single in­vest­ig­a­tion pack­age, cut­ting in­vest­ig­a­tion time from hours to minutes (Mi­crosoft, 2026).

What the log layer has to keep, and for how long

Log­ging an agent's ac­tion is not the same as being able to pro­duce it, and the second is what an audit tests. Four de­cisions de­term­ine whether the trail sur­vives a de­mand for it.

  • Re­ten­tion. Fix it from the longest ob­lig­a­tion reach­ing the es­tate rather than from stor­age cost. CERT-In's Dir­ec­tions of 28 April 2022, No. 20(3)/2022-CERT-In, re­quire all ser­vice pro­viders, in­ter­me­di­ar­ies, data centres, body cor­por­ate and gov­ern­ment or­gan­isa­tions to en­able logs of all their ICT sys­tems and main­tain them se­curely for a rolling period of 180 days. India's Di­gital Per­sonal Data Pro­tec­tion Rules, 2025, im­pose a min­imum one-year re­ten­tion of logs of pro­cessing from 13 May 2027, bind­ing on a pro­cessor as well as on a fi­du­ciary.
  • Loc­a­tion. Dir­ec­tion (iv) of the CERT-In Dir­ec­tions states that the logs shall be main­tained within the In­dian jur­is­dic­tion. CERT-In's own FAQ of May 2022, at ques­tion 35, states that logs may be stored out­side India as long as the ob­lig­a­tion to pro­duce them in reas­on­able time is met. The bind­ing text and the FAQ are in ten­sion. A design should cite both and choose de­lib­er­ately rather than quote whichever suits.
  • Gran­u­lar­ity. The re­cord has to an­swer what was done, on whose be­half, under what policy and why. A log cap­tur­ing the call but not the policy de­cision an­swers one ques­tion of the four.
Zero Trust SecurityVoice of Volge10 •
  • Pro­du­cib­il­ity. Meas­ure the time taken to as­semble a week of one agent's ac­tions. An es­tate that man­ages it in an af­ter­noon will pass an audit. One that needs a pro­ject will not.

Two ob­lig­a­tions at­tach to the in­cid­ent rather than to the log, and both are In­dian. Re­port­able in­cid­ents go to CERT-In within six hours of being no­ticed, and An­nex­ure I of the Dir­ec­tions lists, at item xx, at­tacks or ma­li­cious or sus­pi­cious activ­it­ies af­fect­ing sys­tems, serv­ers, soft­ware and ap­plic­a­tions re­lated to ar­ti­fi­cial in­tel­li­gence and ma­chine learn­ing. In­cid­ents in­volving an AI sys­tem have there­fore been ex­pressly re­port­able in India since 2022. The duty is neither trans­fer­able nor cap­able of being in­dem­ni­fied, per ques­tion 13 of the same FAQ, so it can­not be passed to the plat­form vendor by con­tract.

Audit questionWhere answeredMechanism
Which agent acted?Sign-in / audit logsUnique Agent ID per action
On whose behalf?Sign-in logsDelegated vs own-access flow record
Under what policy?Conditional Access logsPolicy evaluation result, including report-only
What data was touched?Purview audit / DSPM for AIDLP, sensitivity labels, classification
Was it anomalous?Identity ProtectionAgent risk score; Defender verdict

Gap four , revocation that has actually been tested

The last mech­anic is the power to cut an agent's au­thor­ity in­stantly, and it is the one most often as­sumed and least often ex­er­cised. Re­voc­a­tion is the off­board­ing of a di­gital em­ployee, and it has to be as re­li­able as the off­board­ing of a human one , clos­ing step five of the breach pipeline on de­mand rather than at task end.

Be­cause an Entra Agent ID iden­tity is a first-class, man­aged ob­ject, it can be dis­abled, its ac­cess re­voked, and its tokens in­val­id­ated through the same iden­tity-gov­ernance ma­chinery that off­boards a user. Con­di­tional Ac­cess sup­plies the rapid-re­sponse layer: a policy block­ing a named agent iden­tity, or every agent ex­cept an ap­proved set, takes ef­fect at the next token ac­quis­i­tion. A high-risk sig­nal from Iden­tity Pro­tec­tion blocks an agent auto­mat­ic­ally through the con­di­tional-ac­cess policy of sec­tion 4.3 (Mi­crosoft Learn, 2026).

What sep­ar­ates a cred­ible pro­gramme from a paper one is that re­voc­a­tion is tested, not as­sumed. Ex­er­cise the kill-switch against a live agent in a con­trolled win­dow and con­firm three things: the agent loses re­source ac­cess, its tokens stop work­ing, and the ac­tion is logged. An un­tested kill-switch is a hope, and a breach is the one test no one wants to run first.

The four mechanics as a control matrix

A con­trol is demon­strated by a test, and the tests below are what an aud­itor ac­cepts in place of a con­fig­ur­a­tion screen­shot. Each line names the con­trol, the set­ting that im­ple­ments it, the test, and the sig­na­ture of fail­ure.

ControlWhere it is configuredTestSignature of failure
Dedicated agent identityEntra Agent ID blueprint and agent identityTrace one action from tool call to sign-in record and confirm a unique agent identifierThe action resolves to a shared service principal or to a human account
Registry completenessAgent registry in Microsoft Agent 365Reconcile the registry against directory token activity, billing and network egressAn agent with recent activity appears in only one of the four sources
Tool-level least privilegeTool permissions per agentInvoke a tool outside the agent's task scopeThe call succeeds, or fails without a log entry
Delegation constraintSub-agent permission inheritanceReach a parent-forbidden resource through a sub-agentThe chain succeeds, and effective reach exceeds the documented grant
Credential lifetimeJust-in-time, purpose-bound credentialsComplete a task, then re-attempt authentication with the issued credentialThe credential still authenticates after the task has ended
Approval gateConditional Access on a custom security attribute, with a second policy for agent user accountsCreate an agent without the approved attribute value and attempt resource accessAccess succeeds, usually because only one of the two subjects was covered
Risk-based blockConditional Access on the agent risk conditionRaise a test risk signal and confirm the block, in report-only firstNo evaluation record, or an evaluation that allows
Audit sufficiencyEntra sign-in and audit logs, with Purview and Defender integrationReconstruct a week of one agent's actions from logs aloneA gap at what was done, on whose behalf, under what policy, or why
RevocationIdentity governance lifecycle and a rapid-response policyFire the kill-switch against a live agent in a controlled windowAccess persists past the next token acquisition, or the event is unlogged

Sources: Mi­crosoft Learn, Con­di­tional Ac­cess for agents and Se­cure autonom­ous agents with Con­di­tional Ac­cess, 2026; Cloud Se­cur­ity Al­li­ance Agen­tic Trust Frame­work, 2 Feb­ru­ary 2026; OWASP Top 10 for Agen­tic Ap­plic­a­tions, 9 Decem­ber 2025. The test and fail­ure columns are Volge's own for­mu­la­tion of how each con­trol is evid­enced.

Zero Trust SecurityVoice of Volge11 •

Five ways the configuration fails in practice

Most agent-iden­tity in­cid­ents trace to a small set of con­fig­ur­a­tion er­rors, and each is avoid­able by in­spec­tion. They recur be­cause each one is a reas­on­able de­cision taken once and never re­vis­ited.

  • One policy, two sub­jects. A policy scoped to agent iden­tit­ies does not cover an agent's user ac­count, and a policy scoped through the Users as­sign­ment does not reach agent user ac­counts. An ap­proval gate needs a policy for each sub­ject, and es­tates that build one be­lieve they have both.
  • Del­eg­ated ac­cess as a hid­ing place. Pre­fer­ring del­eg­ated ac­cess nar­rows the blast ra­dius and nar­rows the sig­nal with it, be­cause risky activ­ity in on-be­half-of flows is at­trib­uted to the user and the pub­lished agent risk de­tec­tions apply only to autonom­ous agent activ­ity.
  • Re­port-only as a des­tin­a­tion. A policy val­id­ated in re­port-only and never pro­moted is a meas­ure­ment rather than a con­trol. Date the pro­mo­tion, and re­view the list of un­pro­moted policies monthly.
  • Scope set at the blue­print and then for­got­ten. Policy on a blue­print cov­ers every iden­tity in­stan­ti­ated from it, later ones in­cluded, which is the model's strength and the reason a loose blue­print propag­ates si­lently.
  • Pre­view treated as gen­eral avail­ab­il­ity. The agent risk con­di­tion re­mained in pre­view as at Novem­ber 2026, and a con­trol rest­ing on it needs a stated fall­back, usu­ally the ap­proval-at­trib­ute policy.

The implication for the security agenda

Zero Trust for agents is four con­figured mech­an­ics that close the four gaps the breach pipeline opened: least priv­ilege at the tool and cre­den­tial, con­di­tional ac­cess on every token, audit of every ac­tion, and tested re­voc­a­tion. Each is avail­able in the Entra and Agent 365 es­tate today, and each maps onto the Cloud Se­cur­ity Al­li­ance and OWASP frame­works. What re­mains is to se­quence their de­ploy­ment so the found­a­tion pre­cedes the autonomy, which is chapter four.

, Sequencing the controls

reference architecture, rollout, and maturity

The takeaway

The iden­tity con­trols for agents stack into a layered ar­chi­tec­ture, and the lay­ers have a forced order. Iden­tity comes first. Then the re­gistry, then policy and audit and re­voc­a­tion, be­cause each de­pends on the one be­neath it. The rol­lout runs over roughly twelve months, and a ma­tur­ity model lets a team place it­self and define the next in­cre­ment. The single dis­cip­line that sep­ar­ates suc­cess from rogue ex­pos­ure is that the iden­tity found­a­tion pre­cedes the first autonom­ous agent rather than chas­ing it. Build the lay­ers out of order and the breach pipeline re­appears between them.

The reference architecture

The ar­chi­tec­ture below is anchored to the Mi­crosoft es­tate, con­sist­ent with this paper's lens, but the lay­er­ing is port­able. Its value is that each layer is a dis­tinct con­trol that can be con­figured, audited, and op­er­ated on its own while de­pend­ing on the layer under it, which is also why skip­ping a lower layer leaves a gap the breach pipeline runs straight through.

The iden­tity layer. Mi­crosoft Entra Agent ID, as­sign­ing each agent a first-class iden­tity through the blue­print and agent-iden­tity ob­ject model. The found­a­tion; noth­ing above it func­tions without it.

The re­gistry layer. The agent re­gistry in Mi­crosoft Agent 365, in­vent­ory­ing every agent , Mi­crosoft-built, third-party, and autonom­ous , so the es­tate is vis­ible and at­trib­ut­able.

The policy layer. Con­di­tional Ac­cess for agents, eval­u­at­ing every token ac­quis­i­tion against scope, against the ap­proval at­trib­ute and against agent risk, with Iden­tity Pro­tec­tion sup­ply­ing the risk sig­nal.

The audit layer. The Entra sign-in and audit logs, in­teg­rated with Mi­crosoft Pur­view for data clas­si­fic­a­tion and data-loss pre­ven­tion and with Mi­crosoft De­fender for threat de­tec­tion, cap­tur­ing every ac­tion with a unique Agent iden­ti­fier.

The re­sponse layer. Re­voc­a­tion and the kill-switch, through iden­tity-gov­ernance li­fe­cycle con­trols and rapid-re­sponse con­di­tional-ac­cess policies, tested rather than as­sumed.

The ar­chi­tec­tural prin­ciple is that iden­tity gov­ernance is a layer span­ning the es­tate, not a fea­ture em­bed­ded in each agent. That is what lets one con­trol plane gov­ern agents built in-house, by third parties, and on com­pet­ing plat­forms , the multi-vendor real­ity the ref­er­ence flag­ship es­tab­lished.

Mi­crosoft added two Entra ser­vice plans, Con­di­tional Ac­cess for Agents and ID Pro­tec­tion for Agents, to Mi­crosoft Agent 365 and Mi­crosoft 365 E7, with the world­wide rol­lout of that en­ti­tle­ment run­ning from early July to early Au­gust 2026

Zero Trust SecurityVoice of Volge12 •

(Mi­crosoft 365 Mes­sage Cen­ter MC1395007, June 2026). The no­tice gran­ted li­cens­ing rather than avail­ab­il­ity, and the agent-risk con­di­tion it en­titles re­mains in pre­view. Ten­ants ap­ply­ing these con­trols to agents without Agent 365 or E7 now need one of those li­cences; Entra ID P1 and P2 alone no longer carry the agent scen­arios, and Mi­crosoft Learn (2026) states that an Agent 365 li­cence must sit along­side at least Entra ID P1 or Mi­crosoft 365 E3, and along­side Entra ID P2 or Mi­crosoft 365 E5 where the policy is risk-based. The prac­tical read­ing is that an en­ter­prise already li­cens­ing Agent 365 to run its agents, or on E7, owns the en­ti­tle­ment, and its gap is con­fig­ur­a­tion, not pro­cure­ment. An E5 ten­ant without Agent 365 has a li­cens­ing line to close first, and Mi­crosoft pub­lishes both the price and the route. Agent 365 reached gen­eral avail­ab­il­ity on 1 May 2026 at USD 15 per user per month stan­dalone, or in­side Mi­crosoft 365 E7 at USD 99 per user per month, and is sold as an add-on to Mi­crosoft 365 E5, A5 or Busi­ness Premium (Mi­crosoft, May 2026; Mi­crosoft Product Terms, 2026). The li­cence is coun­ted per user who man­ages or spon­sors agents, or whose work agents do, and not per agent, so the line scales with the gov­ern­ing pop­u­la­tion rather than with the agent pop­u­la­tion.

The licensing prerequisites

The con­trols carry spe­cific li­cens­ing pre­requis­ites, and a team should con­firm them be­fore com­mit­ting to a se­quence. Mi­crosoft sets them out as fol­lows (Mi­crosoft Learn, 2026).

CapabilityLicence required
Agent identity platform (create and manage agent identities)Available to all Microsoft Entra customers
Operate agents across Microsoft 365 and enterprise workflowsMicrosoft Agent 365 licence per user
Conditional Access for agentsMicrosoft 365 E7, or a Microsoft Agent 365 licence paired with Entra ID P1 or Microsoft 365 E3 (Entra Conditional Access for Agents service plan)
Risk-based Conditional Access for agents (agent risk condition)As above, with Microsoft Entra ID P2 or Microsoft 365 E5; the agent risk condition is in preview
Identity Protection for agentsMicrosoft Agent 365 or Microsoft 365 E7 (Entra ID Protection for Agents service plan)
Identity Governance for agentsMicrosoft Entra ID P1
Network controls for agentsMicrosoft Entra Internet Access (Entra Suite or separate)

The rollout roadmap

The roadmap se­quences the five lay­ers over roughly twelve months. It is de­lib­er­ately con­ser­vat­ive on agent autonomy and de­lib­er­ately ag­gress­ive on iden­tity found­a­tion , the

in­verse of the fail­ure pat­tern in which agents pro­lif­er­ate be­fore the con­trols exist.

Stage one, the first 60 days , es­tab­lish iden­tity and in­vent­ory. Con­fig­ure Mi­crosoft Entra Agent ID. Au­thor agent blue­prints car­ry­ing baseline se­cur­ity policy. Stand up the agent re­gistry in Agent 365 and on­board every ex­ist­ing agent into it, in­clud­ing the third-party and shadow agents sur­faced in the pro­cess. The ob­ject­ive is a com­plete, at­trib­ut­able in­vent­ory and a first-class iden­tity for every agent. Past this stage, no autonom­ous agent op­er­ates out­side the re­gistry.

What stage one actually involves

Stage one is dis­cov­ery, and it is un­glam­or­ous. Most of the ef­fort goes into find­ing agents nobody in­ten­ded to hide, and the work runs bet­ter as a time-boxed sweep with a named owner than as a re­quest to the busi­ness.

Ex­pect three kinds of sur­prise. Agents built in­side a sanc­tioned plat­form by people out­side the cent­ral team, which the re­gistry sur­faces dir­ectly. Agents built on an un­sanc­tioned plat­form and paid for on a card, which the pro­cure­ment and billing re­cord sur­faces. And auto­ma­tions that are agents in sub­stance but were never called one, usu­ally a script hold­ing a cre­den­tial and call­ing a model en­d­point, which only net­work egress sur­faces. The third group is the smal­lest and car­ries the largest ex­pos­ure per item.

De­cide the dis­pos­i­tion of each dis­covered agent at the point of dis­cov­ery, be­cause a list without de­cisions be­comes a list nobody re­vis­its. Four dis­pos­i­tions suf­fice. On­board it to a first-class iden­tity and the re­gistry. Re­tire it, where the task it served has ended. Re­place it, where a sanc­tioned plat­form already does the same work under gov­ernance. Or grant it a dated ex­cep­tion with a named owner and a re­view date, which should be rare and vis­ibly coun­ted.

State the stage-one gate be­fore the sweep be­gins. Past this stage, no autonom­ous agent op­er­ates out­side the re­gistry, and the ex­cep­tion list is re­por­ted monthly with the age of each entry. A gate de­clared af­ter­wards is a pref­er­ence.

Stage two, months two to five , en­force policy and least priv­ilege. De­ploy the two ca­non­ical con­di­tional-ac­cess policies in re­port-only mode: the ap­proval-at­trib­ute block and the high-risk block. Re­view im­pact, then pro­mote to en­force­ment. Con­vert stand­ing agent cre­den­tials to just-in-time, pur­pose-bound cre­den­tials. Scope tool per­mis­sions per agent and con­strain del­eg­a­tion in­her­it­ance. The ob­ject­ive is least priv­ilege at the tool and the cre­den­tial, with policy eval­u­at­ing every token.

The meas­ure of a good cutover is the count of un­planned re­versals, and it should sit close to zero. A pro­gramme that has

Zero Trust SecurityVoice of Volge13 •

never re­versed an en­force­ment has prob­ably never en­forced at scale.

Cutover without breaking the workloads

The risk in stages two and three is not the con­trol. It is the out­age caused by ap­ply­ing the con­trol to a live es­tate, and that risk is what makes a se­cur­ity team defer the work in­def­in­itely. Four habits con­tain it.

  • Re­port-only first, al­ways, and read the eval­u­ation re­cords rather than the sum­mary. The pur­pose is to find the agent nobody re­membered, not to con­firm that the policy logic is sound.
  • En­force by co­hort rather than across the ten­ant. Start with a co­hort whose own­ers are avail­able that week, and keep it small enough that a roll­back is a de­cision rather than an in­cid­ent.
  • Pair every en­force­ment with a named roll­back and the per­son who can ex­ecute it. An en­force­ment with no roll­back owner will be re­versed by some­body guess­ing under pres­sure.
  • Keep a stand­ing ex­cep­tion path with a short ex­piry. Teams route around a con­trol that of­fers no le­git­im­ate es­cape, and a dated ex­cep­tion is vis­ible where a work­around is not.

Stage three, months five to nine, in­stru­ment audit and in­teg­rate. Con­firm that every au­then­tic­a­tion and token ac­quis­i­tion is logged with a unique Agent iden­ti­fier. Con­firm the same for every tool in­voc­a­tion and policy de­cision. In­teg­rate Entra logs with Mi­crosoft Pur­view for data clas­si­fic­a­tion and data-loss pre­ven­tion and with Mi­crosoft De­fender for threat de­tec­tion. Val­id­ate that an agent ac­tion as­sembles into a single in­vest­ig­a­tion pack­age. The ob­ject­ive is an audit trail that would pass the agent-be­ha­viour audit 84 per cent of or­gan­isa­tions doubt they could pass.

Stage four, months nine to twelve, test re­voc­a­tion and op­er­a­tion­al­ise. Ex­er­cise the kill-switch against a live agent in a con­trolled win­dow, con­firm­ing that ac­cess is lost, that tokens are in­val­id­ated and that the event is logged. Fold agent off­board­ing into the iden­tity-gov­ernance li­fe­cycle so de­com­mis­sion­ing is routine rather than ex­cep­tional. The ob­ject­ive is a tested re­voc­a­tion cap­ab­il­ity and a steady-state op­er­at­ing rhythm for the es­tate.

What each stage has to evidence

A stage is com­plete when it can pro­duce evid­ence, not when the con­fig­ur­a­tion is saved. Four meas­ures and four arte­facts settle it.

StageMeasureArtefact
One, identity and inventoryShare of agents holding a dedicated identity, and share present in the registry, against the reconciled populationA reconciliation record naming every difference between the four discovery sources, and its resolution
Two, policy and least privilegeShare of agents scoped at the tool, share on just-in-time credentials, and the count of policies promoted out of report-onlyBoth canonical policies in enforcement, each with the dated report-only impact review behind it
Three, audit and integrationSampled share of agent actions reconstructable from logs aloneOne completed investigation package assembled from Entra, Purview and Defender for a single agent action
Four, revocation and operationsTime from decision to confirmed loss of access, measured in a drillA dated revocation drill report per agent class, and agent offboarding written into the joiner-mover-leaver process

Sample the third meas­ure rather than de­riv­ing it from con­fig­ur­a­tion. At­tri­bu­tion com­plete­ness de­grades as the es­tate grows, and con­fig­ur­a­tion will re­port a pos­ture the logs do not sup­port.

What the programme reports, and to whom

An agent es­tate needs a stand­ing re­port, be­cause the ex­pos­ure moves faster than an an­nual re­view. Keep it short, and keep it the same each month.

  • Pop­u­la­tion and change. Agents by ac­cess pat­tern, ad­di­tions and de­com­mis­sion­ings, and the count dis­covered out­side the re­gistry.
  • Scope drift. Agents whose con­sen­ted scopes widened in the period, with the ap­proval be­hind each.
  • Policy cov­er­age. Agents out­side the ap­proval gate, and policies still sit­ting in re­port-only.
  • At­tri­bu­tion. The sampled share of ac­tions re­con­struct­able from logs, with the sample size stated.
  • Re­voc­a­tion read­i­ness. Agent classes whose drill is older than a quarter.

The audi­ence is the chief in­form­a­tion se­cur­ity of­ficer monthly and the audit com­mit­tee quarterly. The last line is the one that de­cays without any­one de­cid­ing to let it, which is why it is re­por­ted rather than as­sumed.

The identity maturity model

The model lets a team place the en­ter­prise and define the next in­cre­ment. It is the dia­gnostic coun­ter­part to the roadmap and aligns with the In­tern-to-Prin­cipal pro­gres­sion of the Cloud Se­cur­ity Al­li­ance frame­work, in which agents earn autonomy through demon­strated trust­wor­thi­ness (Cloud Se­cur­ity Al­li­ance, 2026). Read the levels against

Zero Trust SecurityVoice of Volge14 •

chapter one: Level 0 is the breach pipeline run­ning un­broken; each higher level breaks one more step.

LevelIdentity stateAccess controlAudit and revocationTypical exposure
0 , InvisibleAgents use shared keys or human credentials; no inventoryNone scoped to agentsNo agent-specific logging; no kill-switchRogue exposure; cannot pass audit
1 , IdentifiedEvery agent has a first-class Entra Agent ID identity; registered in Agent 365Inherited or broadAgent actions logged with unique IDVisible but over-permissioned
2 , ScopedFirst-class identity; delegated vs own-access classifiedConditional Access enforced; least privilege at tool and credentialFull sign-in and audit logging; revocation definedControlled; audit-ready
3 , VerifiedIdentity plus continuous risk evaluationRisk-adaptive Conditional Access; JIT credentials estate-widePurview and Defender integrated; kill-switch testedLow; continuously verified
4 , Governed estateEstate-wide identity spanning Microsoft and third-party agentsPolicy spans vendors; delegation constrained end-to-endUnified investigation; routine offboardingMinimal; demonstrable to auditor

Most en­ter­prises sit between Level 0 and Level 1 as at late 2026, con­sist­ent with the find­ing that only 21 per cent main­tain a real-time re­gistry of their agents (Cloud Se­cur­ity Al­li­ance, Feb­ru­ary 2026). The audit-ready threshold is Level 2.

The implication for the security agenda

The con­trols are a se­quence and not a menu. That se­quence is iden­tity, then re­gistry and policy, then audit and re­voc­a­tion. The en­ter­prise that de­ploys them in that order, with the iden­tity found­a­tion ahead of the first autonom­ous agent, reaches the audit-ready Level 2 within months. The en­ter­prise that de­ploys agents first and con­trols later re­builds the rogue ex­pos­ure of chapter one between its half-built lay­ers. The order is the dis­cip­line.

The con­trol plane sits in the same po­s­i­tion. Mi­crosoft states that Agent 365 pro­cesses and stores the traces for agents built across Mi­crosoft ser­vices, that it "cur­rently stores and pro­cesses the data at the ten­ant geo home loc­a­tion", and that macro re­gion geo­graphy and en­vir­on­ment-level re­gion se­lec­tion do not apply to it (Mi­crosoft Learn, page up­dated 7 Oc­to­ber 2026). Agent 365 reached gen­eral avail­ab­il­ity for the com­mer­cial seg­ment on 1 May 2026. Entra-backed agent iden­tity is de­livered through the Agent 365 soft­ware

de­vel­op­ment kit, so on the pub­lished evid­ence it in­her­its that ab­sence of a re­gional dial.

, Identity in the Gulf and India

The identity layer is the one layer Microsoft does not commit to keeping in region, and it says so in its own documentation. The regulator, meanwhile, requires nothing of non-human identity anywhere in the Gulf or India, which is a verified finding rather than a research gap. Those two facts frame the regional position. The control plane is less localised than the data it governs, and the operator has to write the rule it will later be audited against.

Every res­id­ency ar­gu­ment an en­ter­prise makes about its agents has a hole in the middle of it, and the hole is iden­tity. Mi­crosoft's Co­pi­lot Stu­dio and Dy­nam­ics 365 res­id­ency doc­u­ment­a­tion both name Mi­crosoft Entra ID as a ser­vice that might store data glob­ally, and Entra mul­ti­factor au­then­tic­a­tion data as data that may be stored glob­ally, as ex­press ex­cep­tions to a ten­ant's chosen geo­graphy (Mi­crosoft Learn, pages up­dated 30 April and 7 Oc­to­ber 2026). Mi­crosoft's Mi­crosoft 365 doc­u­ment­a­tion does not carry Entra res­id­ency at all, dir­ect­ing in­stead to a sep­ar­ate Entra data map. The iden­tity layer is com­mit­ted to no re­gion, and no add-on changes that.

The con­sequence is pre­cise. Agent traces are the audit re­cord of what the agents did, and they fol­low the ten­ant rather than the work­load. India, Qatar and the United Arab Emir­ates are Local Re­gion Geo­graph­ies, so a ten­ant in any of the three has a home geo­graphy in coun­try, and the po­s­i­tion is tol­er­able. Saudi Ar­a­bia, Kuwait, Oman and Bahrain are not Local Re­gion Geo­graph­ies, and Mi­crosoft places their ten­ants in the cat­egory it calls All Other Geo­graph­ies, where cus­tomer data "isn't com­mit­ted to reside in any par­tic­u­lar data­cen­ter" and the loc­a­tion "is sub­ject to change without no­tice" (Mi­crosoft Learn, page up­dated 27 Oc­to­ber 2026). For those four, the gov­ernance evid­ence for the agent es­tate sits wherever Mi­crosoft puts the ten­ant.

JurisdictionMicrosoft 365 Local Region GeographyAgent 365 trace locationEntra ID and multifactor authentication data
India, Qatar, United Arab EmiratesYes, with durable commitments on the six core workloadsTenant home geography, in countryMay be stored globally
Saudi Arabia, Kuwait, Oman, BahrainNo, and not eligible for the Advanced Data Residency add-onTenant home geography, with no commitment to a locationMay be stored globally

Sources: Mi­crosoft Learn data res­id­ency and Agent 365 doc­u­ment­a­tion, pages up­dated between 7 and 27 Oc­to­ber 2026, veri­fied 8 Novem­ber 2026. Com­pil­a­tion by Volge.

Zero Trust SecurityVoice of Volge15 •

What Advanced Data Residency buys, and what it does not

The paid res­id­ency add-on is bought for the wrong reason more often than not. Ad­vanced Data Res­id­ency is avail­able only where a ten­ant's de­fault geo­graphy is a Local Re­gion Geo­graphy, which among the jur­is­dic­tions here means India, Qatar or the UAE. The six core work­loads, being Ex­change On­line, SharePoint, OneDrive, Teams, Co­pi­lot and Co­pi­lot Chat, already carry a Product Terms com­mit­ment for an eli­gible ten­ant. What the add-on uniquely adds is a dur­able com­mit­ment over De­fender for Of­fice Plan 1, Ex­change On­line Pro­tec­tion, Of­fice for the web, Viva Con­nec­tions and six named Pur­view ser­vices, those being Audit Stand­ard, Audit Premium, Data Li­fe­cycle Man­age­ment, Data Loss Pre­ven­tion, In­form­a­tion Bar­ri­ers and In­form­a­tion Pro­tec­tion (Mi­crosoft Learn, page up­dated 27 Oc­to­ber 2026).

Read that list against an agent-gov­ernance pro­gramme and the value be­comes clearer than the sales fram­ing. The Pur­view audit and in­form­a­tion-pro­tec­tion ser­vices carry much of the agent audit trail, and they are the part of the es­tate the add-on genu­inely loc­al­ises. The cost struc­ture is the catch. Mi­crosoft re­quires the cus­tomer to cover every paid li­cence in the ten­ant with an add-on li­cence, so there is no par­tial ad­op­tion. And noth­ing in the add-on reaches Entra, Agent 365 or the in­fer­ence path.

Three de­faults sur­vive every res­id­ency pur­chase and be­long in the design doc­u­ment. Pre­view and pre-re­lease ser­vices typ­ic­ally store cus­tomer data in the United States and may store it glob­ally. His­tor­ical agent activ­ity data in Co­pi­lot Stu­dio, con­ver­sa­tion logs in­cluded, is stored in the re­gion of the end user's Ex­change mail­box rather than in­side the Power Plat­form bound­ary, which moves agent logs from one res­id­ency re­gime into an­other. And every Gulf and In­dian ten­ant's Mi­crosoft Forms data sits in the United States (Mi­crosoft Learn, page up­dated 8 Oc­to­ber 2026).

The regulator requires nothing of machine identity, and that is the finding

No in­stru­ment in Saudi Ar­a­bia, the UAE, Qatar, Kuwait, Oman, Bahrain or India reg­u­lates non-human or ma­chine iden­tity as a cat­egory. No in­stru­ment read defines a ser­vice ac­count, an ap­plic­a­tion ac­count, a sys­tem ac­count, a ma­chine iden­tity or a non-human iden­tity, and no ob­lig­a­tion any­where in the seven jur­is­dic­tions re­quires an op­er­ator to in­vent­ory, ro­tate, at­test or de-pro­vi­sion ser­vice ac­counts, work­load iden­tit­ies or agent cre­den­tials. That is a veri­fied neg­at­ive, es­tab­lished by ex­haust­ive text search of the primary doc­u­ments, and it should be stated rather than apo­lo­gised for. Iden­tity

ob­lig­a­tions across the re­gion are writ­ten for human users and, at their most ad­vanced, for shared priv­ileged ac­counts.

The nearest hooks are worth know­ing pre­cisely, be­cause they are what a su­per­visor reaches for. The best of them is In­dian and it is about hard­ware se­cur­ity mod­ules. Para­graphs 70(3) to (5) of the Re­serve Bank of India's Mas­ter Dir­ec­tion on Di­gital Pay­ment Se­cur­ity Con­trols, 2021, re­quire ac­cess con­trol lists main­tained sep­ar­ately for each in­di­vidual ap­plic­a­tion to en­sure ap­plic­a­tion-level isol­a­tion, with all HSM ac­cess man­aged and mon­itored through a priv­ileged iden­tity and ac­cess man­age­ment solu­tion. Per-ap­plic­a­tion isol­a­tion of en­ti­tle­ments is, in sub­stance, an ap­plic­a­tion-iden­tity re­quire­ment, though the in­stru­ment never uses that lan­guage.

  • RBI Com­mer­cial Banks Dir­ec­tions, 2026, para­graph 107. Re­quires pro­tec­tion of user ac­cess cre­den­tials such as logon iden­ti­fi­ers, au­then­tic­a­tion in­form­a­tion and tokens against leak­age and at­tack. The only cre­den­tial-ma­ter­ial ob­lig­a­tion in cur­rent RBI law.
  • RBI IT Gov­ernance re­quire­ments of 2023, sec­tion 17(b), car­ried into the 2026 Dir­ec­tions at para­graphs 141 to 142. Re­quires ap­plic­a­tion-to-ap­plic­a­tion data trans­fer to be se­curely auto­mated with an ap­pro­pri­ate au­then­tic­a­tion mech­an­ism and audit trails. The only clause found any­where in the seven jur­is­dic­tions that ex­pli­citly au­then­tic­ates a ma­chine-to-ma­chine flow.
  • SAMA Cyber Se­cur­ity Frame­work, con­trol 3.3.5 item 4.f.4. Ad­dresses the use of non-per­sonal priv­ileged ac­counts, with lim­it­a­tion and mon­it­or­ing, pass­word con­fid­en­ti­al­ity, and pass­word change at the end of each ses­sion. The only Gulf con­trol nam­ing a non-human ac­count class, and it is framed as pass­word hy­giene.
  • SEBI CSCRF stand­ard PR.AA.S5. Re­quires re­view of user ac­cess rights, del­eg­ated ac­cess and un­used tokens, quarterly for mar­ket in­fra­struc­ture in­sti­tu­tions and qual­i­fied en­tit­ies and half-yearly for the rest. The only phrase in the frame­work reach­ing bey­ond human ac­counts.
  • Cent­ral Bank of Kuwait Cyber and Op­er­a­tional Re­si­li­ence Frame­work, launched 3 Decem­ber 2025. Car­ries 876 con­trols against 291 in the 2020 frame­work, with 211 in the third-party risk baseline alone, whose higher ma­tur­ity at­trib­utes name just-in-time ac­cess con­trols and real-time de­pend­ency map­ping. The nearest thing to an agent-ready su­per­vis­ory text in the Gulf, and it ar­rives through third-party risk rather than through iden­tity. Struc­ture per RSM Al­bazie, the frame­work it­self re­turn­ing no ma­chine-read­able text.

Saudi Ar­a­bia's iden­tity as­sets, Na­fath and Tawakkalna, are human-iden­tity ser­vices and should not be cited as a

Zero Trust SecurityVoice of Volge16 •

pre­ced­ent for agent iden­tity. The only reg­u­lat­ory text in the Gulf that re­cog­nises an agent as a thing to be se­cured is the Na­tional Cy­ber­se­cur­ity Au­thor­ity's draft AI Cy­ber­se­cur­ity Guidelines, AICG-1:2026, con­sul­ted from 5 July to 5 Au­gust 2026, whose gloss­ary defines AI Agent, Agen­tic AI and Tool or Plu­gin and whose ab­bre­vi­ation list car­ries Model Con­text Pro­tocol. It is a draft, it is non-bind­ing, and no final in­stru­ment had been pub­lished as at 8 Novem­ber 2026.

The one non-human legal con­struct in the re­gion sits out­side cyber reg­u­la­tion al­to­gether. The UAE's Elec­tronic Trans­ac­tions and Trust Ser­vices law, with Cab­inet Res­ol­u­tion No. 28 of 2023, has the Tele­com­mu­nic­a­tions and Di­gital Gov­ern­ment Reg­u­lat­ory Au­thor­ity li­cens­ing trust ser­vice pro­viders that cre­ate, val­id­ate and pre­serve elec­tronic sig­na­tures, elec­tronic seals and di­gital cer­ti­fic­a­tion. An eSeal is a legal per­son's seal rather than a nat­ural per­son's sig­na­ture, which makes it the closest stat­utory an­chor for a ma­chine cre­den­tial any­where in the Gulf. The in­stru­ment text was not read, so the claim here is con­fined to what the li­cens­ing frame­work es­tab­lishes.

What an operator does when there is no rule

The ab­sence of a rule shifts the work rather than re­mov­ing it, be­cause three ob­lig­a­tions reach an agent es­tate without nam­ing it. Ac­count­ab­il­ity for an AI sys­tem's out­put is al­loc­ated in India. Data pro­tec­tion law ap­plies to whatever the agent pro­cesses. And third-party ac­count­ab­il­ity rules apply to the plat­form the agent runs on. Design to those three and the even­tual ma­chine-iden­tity rule will find the con­trols already in place.

The ac­count­ab­il­ity point is the sharpest of the three. Reg­u­la­tion 16C of the SEBI (In­ter­me­di­ar­ies) (Amend­ment) Reg­u­la­tions, 2025, in force since early Feb­ru­ary 2025, places sole re­spons­ib­il­ity for the out­put of an ar­ti­fi­cial-in­tel­li­gence or ma­chine-learn­ing tool on the reg­u­lated per­son using it, in-house or pro­cured, ir­re­spect­ive of the scale and scen­ario of ad­op­tion. It binds In­dian se­cur­it­ies-mar­ket in­ter­me­di­ar­ies only. It is non­ethe­less the only pro­vi­sion in the seven jur­is­dic­tions that al­loc­ates li­ab­il­ity for an AI sys­tem's out­put, and it settles the ques­tion a board asks first. The firm that de­ploys the agent an­swers for what the agent does, and pro­cure­ment does not trans­fer that.

SEBI's Cy­ber­se­cur­ity and Cyber Re­si­li­ence Frame­work states the third-party half in terms. Reg­u­lated en­tit­ies are solely ac­count­able for all as­pects of third-party ser­vices, in­clud­ing con­fid­en­ti­al­ity, in­teg­rity, avail­ab­il­ity, non-re­pu­di­ation and the se­cur­ity of their data and logs, and are re­spons­ible for any vi­ol­a­tion. Ap­plied to an agent es­tate run­ning on a vendor's con­trol plane, that makes the audit trail the firm's ob­lig­a­tion even where the vendor holds it, and it turns the Agent 365 trace loc­a­tion into the firm's prob­lem rather than Mi­crosoft's.

The data-pro­tec­tion half car­ries dates. India's Di­gital Per­sonal Data Pro­tec­tion Rules, 2025, no­ti­fied on 13 Novem­ber 2025 as G.S.R. 846(E), com­mence in phases. Con­sent man­agers come into force on 13 Novem­ber 2026, and the sub­stant­ive ob­lig­a­tions on 13 May 2027, cov­er­ing no­tice, se­cur­ity safe­guards, breach in­tim­a­tion, eras­ure, chil­dren's data, Sig­ni­fic­ant Data Fi­du­ciary du­ties, data prin­cipal rights and cross-bor­der trans­fer. Until that date sec­tion 43A of the In­form­a­tion Tech­no­logy Act and the 2011 SPDI Rules re­main the op­er­at­ive In­dian data se­cur­ity pro­vi­sions. Rules 6(1)(e) and 8(3) im­pose a min­imum one-year re­ten­tion of per­sonal data and of logs of pro­cessing, bind­ing on a pro­cessor en­gaged by a fi­du­ciary as well, which cuts against a de­le­tion-by-de­fault log­ging policy for agent activ­ity. Rule 6(1) re­quires en­cryp­tion or ob­fus­ca­tion or mask­ing, ac­cess con­trol, and vis­ib­il­ity of ac­cess through logs, mon­it­or­ing and re­view.

The UAE po­s­i­tion car­ries a gap worth stat­ing plainly. Fed­eral De­cree-Law No. 45 of 2021, the Per­sonal Data Pro­tec­tion Law, still has no Ex­ec­ut­ive Reg­u­la­tions, nearly five years past its stat­utory dead­line. It there­fore has no com­pli­ance com­mence­ment date, no breach no­ti­fic­a­tion dead­line, no pen­alty sched­ule and no func­tion­ing reg­u­lator, and any claim that the UAE im­poses a 72-hour breach no­ti­fic­a­tion is un­sup­por­ted. The bind­ing AI-ad­ja­cent in­stru­ment in the Emir­ates is DIFC Data Pro­tec­tion Reg­u­la­tions Reg­u­la­tion 10, in force 1 Septem­ber 2023, which al­loc­ates ac­count­ab­il­ity for an autonom­ous or semi-autonom­ous sys­tem between the De­ployer, who is ac­count­able, and the Op­er­ator, who is the tech­nical pro­vider act­ing on the de­ployer's in­struc­tions. It al­loc­ates by role rather than by iden­tity, and it binds in­side the Dubai In­ter­na­tional Fin­an­cial Centre alone.

India's Zero Trust mandate, and what its softening teaches

India wrote the re­gion's only dated Zero Trust re­quire­ment and then with­drew its hard edge, and an iden­tity team should read the arc rather than the en­d­point. SEBI's Cy­ber­se­cur­ity and Cyber Re­si­li­ence Frame­work, cir­cu­lar SEBI/HO/ITD-1/ITD_CSC_EXT/P/CIR/2024/113 of 20 Au­gust 2024, re­quired under stand­ards PR.AA.S4 and PR.AA.S5 a zero-trust se­cur­ity model in which ac­cess to crit­ical sys­tems, from in­side or out­side the en­tity's net­work, is denied by de­fault and al­lowed only after proper au­then­tic­a­tion and au­thor­isa­tion. Tech­nical Cla­ri­fic­a­tions cir­cu­lar SEBI/HO/ITD-1/ITD_CSC_EXT/P/CIR/2025/119 of 28 Au­gust 2025 re­placed it with a re­quire­ment to im­ple­ment sug­ges­ted strategies and meth­od­o­lo­gies such as zero-trust net­works, seg­ment­a­tion, no single point of fail­ure and high avail­ab­il­ity, ap­proved by the en­tity's IT com­mit­tee.

Zero Trust SecurityVoice of Volge17 •

A de­fault-deny man­date las­ted twelve months and be­came a com­mit­tee-ap­proved menu. The les­son is not that the reg­u­lator re­treated from Zero Trust. A de­fault-deny rule writ­ten against an es­tate of human users and leg­acy ap­plic­a­tions proved un­en­force­able at the pace the cir­cu­lar set, and the de­cision moved to the firm's own com­mit­tee. An agent pop­u­la­tion is the one place where de­fault-deny is straight­for­ward, be­cause every agent is new, every agent has a defin­able task, and noth­ing leg­acy de­pends on its stand­ing ac­cess. The right read­ing is to apply to agents the pos­ture the reg­u­lator could not im­pose on everything.

The rest of the frame­work stayed de­mand­ing and is in force, every com­pli­ance date hav­ing passed. It re­quires a board-ap­proved list of crit­ical sys­tems, a doc­u­mented au­then­tic­a­tion and ac­cess policy with log col­lec­tion and re­ten­tion, net­work seg­ment­a­tion, se­cur­ity mon­it­or­ing through a se­cur­ity op­er­a­tions centre, and peri­odic audit by a CERT-In em­pan­elled audit­ing or­gan­isa­tion. Ap­plic­a­tion pro­gram­ming in­ter­face se­cur­ity must be im­ple­men­ted with rate lim­it­ing, throt­tling and proper au­then­tic­a­tion and au­thor­isa­tion mech­an­isms, which is the only ex­press API se­cur­ity man­date in In­dian fin­an­cial-sec­tor reg­u­la­tion. It does not say whose iden­tity au­then­tic­ates, which is pre­cisely the gap an agent falls into. Re­cov­ery tar­gets are a two-hour re­cov­ery time ob­ject­ive for crit­ical op­er­a­tions and a 15-minute re­cov­ery point ob­ject­ive for crit­ical sys­tems. Data loc­al­isa­tion under stand­ard PR.DS.S2 has been in abey­ance since 31 Decem­ber 2024 and re­mains so.

One struc­tural fea­ture of the 28 Au­gust 2025 cla­ri­fic­a­tions de­serves at­ten­tion from any group op­er­at­ing across In­dian reg­u­lat­ors. The Prin­ciple of Ex­clus­iv­ity con­fines the frame­work to sys­tems used ex­clus­ively for SEBI-reg­u­lated activ­it­ies. The Prin­ciple of Equi­val­ence deems com­pli­ant any con­trol with an equi­val­ent in the primary reg­u­lator's frame­work, provided that frame­work is ad­hered to. For a bank that is also a se­cur­it­ies in­ter­me­di­ary, those two prin­ciples de­cide which su­per­visor's iden­tity con­trols gov­ern a shared agent plat­form, and the ques­tion is bet­ter answered be­fore the plat­form is built than dur­ing an in­spec­tion.

The demand that makes this urgent

Gulf agent ad­op­tion is a stated gov­ern­ment com­mit­ment with named ac­count­ab­il­ity, which com­presses the timetable for the iden­tity work. The UAE Cab­inet, on 23 April 2026, ap­proved a frame­work to move half of fed­eral gov­ern­ment sec­tors, ser­vices and op­er­a­tions to agen­tic AI for autonom­ous ex­e­cu­tion and de­cision-mak­ing within two years, with min­is­ters, dir­ect­ors gen­eral and en­tit­ies as­sessed on ad­op­tion speed. Over­sight sits with Sheikh Man­sour bin Zayed, and the taskforce is chaired by Mo­hammad Al Ger­gawi. The status is

an an­nounced frame­work with ap­proved Cab­inet res­ol­u­tions and no pub­lished in­stru­ment text. The same meet­ing ap­proved a Gov­ern­ment Ser­vices Data Shar­ing Policy framed as col­lect once and use se­curely across en­tit­ies, and a fed­eral guide en­cour­aging reuse of di­gital iden­tity cap­ab­il­it­ies. An iden­tity team in­side or sup­ply­ing a fed­eral en­tity should as­sume the agents ar­rive be­fore the con­trols, un­less the con­trols are se­quenced first.

Ac­cred­it­a­tion is the other timetable. A pro­vider of­fer­ing cloud ser­vices to Dubai gov­ern­ment and semi-gov­ern­ment en­tit­ies must hold the Dubai Elec­tronic Se­cur­ity Cen­ter's Cloud Ser­vice Pro­vider Se­cur­ity Stand­ard cer­ti­fic­a­tion, which DESC de­scribes as man­dat­ory in its own words, with an­nual sur­veil­lance audits and re­cer­ti­fic­a­tion every three years. In Saudi Ar­a­bia a pro­vider serving gov­ern­ment must hold CST re­gis­tra­tion at Class A or above, and the class de­term­ines which data clas­si­fic­a­tions it may touch, with Class A lim­ited to pub­lic data. Kuwait re­quires a CITRA cloud li­cence with an in-Kuwait data centre to serve the pub­lic sec­tor, and Oman re­quires a Tele­com­mu­nic­a­tions Reg­u­lat­ory Au­thor­ity au­thor­isa­tion be­fore con­tract­ing. Those four are the only self-de­clared man­dat­ory ac­cred­it­a­tion gates across the seven jur­is­dic­tions, and not one of them asks a ques­tion about agent iden­tity.

The implication for the security agenda

The re­gional po­s­i­tion does not change the con­trol set. It changes where the evid­ence lives and who writes the rule. Treat Entra and Agent 365 as the parts of the es­tate that are not in re­gion, say so in the design doc­u­ment, and stop present­ing a stor­age com­mit­ment as a pro­cessing one. For a ten­ant out­side a Local Re­gion Geo­graphy, de­cide de­lib­er­ately whether ten­ant-bound agent traces are ac­cept­able, be­cause that de­cision can­not be con­figured away. And where no in­stru­ment defines a ma­chine iden­tity, adopt the con­trol, map it to the OWASP and Cloud Se­cur­ity Al­li­ance ref­er­ences an aud­itor re­cog­nises, and re­cord the de­cision through the en­ter­prise's own gov­ernance. Li­ab­il­ity for the agent's out­put already sits with the de­ploy­ing firm in the one jur­is­dic­tion that has ad­dressed it, and the su­per­vis­ory texts are mov­ing that way every­where else.

, The identity leadership agenda

The ques­tion that titles this paper , who is ac­cess­ing your data , has a de­term­in­ate an­swer in any well-gov­erned en­ter­prise and a fright­en­ing one in every other. The evid­ence as­sembled here re­solves into a few clear con­clu­sions for the chief in­form­a­tion se­cur­ity of­ficer and the iden­tity lead­er­ship.

The con­trols to break it exist and are gen­er­ally avail­able. Mi­crosoft Entra Agent ID reached gen­eral avail­ab­il­ity in April

Zero Trust SecurityVoice of Volge18 •

2026 and Mi­crosoft Agent 365 on 1 May 2026 (Mi­crosoft, 2026). The con­straint is not avail­ab­il­ity but con­fig­ur­a­tion and se­quence. Eighty-four per cent of or­gan­isa­tions doubt they could pass an audit of agent be­ha­viour, and the path to clos­ing that gap is the staged rol­lout of chapter four (Cloud Se­cur­ity Al­li­ance, Feb­ru­ary 2026).

The pop­u­la­tion reach­ing cor­por­ate data is pre­dom­in­antly non-human, out­num­ber­ing human users by 109 to one in the en­ter­prises sur­veyed and climb­ing (Palo Alto Net­works, 2026). Agents are its fast­est-grow­ing and least-gov­erned part, and they dif­fer from prior ma­chine iden­tit­ies in that they reason and del­eg­ate, and in that they per­sist. The ex­pos­ure they cre­ate is struc­tural, a five-step pipeline in which every step is a skipped iden­tity con­trol, and it is already pro­du­cing breaches. Break the pipeline at any step and the breach does not occur.

Six ques­tions be­long on the iden­tity lead­er­ship agenda. First, can the en­ter­prise pro­duce a com­plete in­vent­ory of its agents, in­clud­ing third-party and shadow agents, and does every agent hold a first-class iden­tity. Second, is each agent cor­rectly clas­si­fied as del­eg­ated-ac­cess or own-ac­cess, and is own-ac­cess re­served for genu­inely autonom­ous agents scoped from first prin­ciples. Third, is least priv­ilege en­forced at the tool call and the cre­den­tial, with just-in-time ex­piry, or is the agent scoped once at de­ploy­ment and then for­got­ten. Fourth, does Con­di­tional Ac­cess eval­u­ate every token an agent ac­quires, and is the audit trail enough to pass an agent-be­ha­viour audit. Fifth, has the kill-switch been fired against a live agent, or is re­voc­a­tion still an as­sump­tion. Sixth, is the res­id­ency po­s­i­tion writ­ten layer by layer, with Entra and the Agent 365 traces named as the parts that are not in re­gion, and has the firm de­cided who an­swers for the agent's out­put.

What the audit com­mit­tee should see, quarterly, is nar­rower than what the se­cur­ity team tracks. Five lines carry it. The agent pop­u­la­tion by ac­cess pat­tern, with the count found out­side the re­gistry. The share of agents on just-in-time cre­den­tials. The sampled share of agent ac­tions re­con­struct­able from logs alone. The date of the last re­voc­a­tion drill, per agent class. And the res­id­ency po­s­i­tion layer by layer, with the parts that are not in re­gion named. Those five an­swer the ques­tion an aud­itor asks, which is whether the en­ter­prise can demon­strate con­trol rather than de­scribe it.

The con­trols are con­fig­ur­a­tion and se­quence, and both sit in­side the en­ter­prise's own gift. The li­cences are largely already owned, the plat­form de­cision was settled by ex­ist­ing Mi­crosoft and Entra com­mit­ments, and the two frame­works an aud­itor will ref­er­ence were pub­lished within eight weeks of each other. What re­mains is to build the in­vent­ory by re­con­cili­ation rather than by in­vit­a­tion, to scope at the tool and the cre­den­tial rather than at the agent, to pro­mote every policy out of re­port-only on a dated plan, and to fire the kill-switch be­fore an in­cid­ent fires it for the team. An es­tate that has done those four can an­swer the ques­tion that titles this paper. One that has not will an­swer it dur­ing an in­vest­ig­a­tion.

Zero Trust SecurityVoice of Volge19 •

Appendices

Appendix A , Agent identity controls and capability matrix

The fol­low­ing maps each iden­tity con­trol to its Mi­crosoft cap­ab­il­ity and its frame­work ref­er­ence, as at Novem­ber 2026.

ControlMicrosoft capabilityFramework referenceAvailability
First-class agent identityEntra Agent ID (agent identity blueprint, agent identity, optional agent's user account)CSA ATF element 1 (Identity); OWASP excessive agencyGA April 2026
Agent inventory / registryAgent registry in Microsoft Agent 365CSA ATF element 1 (Identity)GA 1 May 2026
Third-party agent onboardingEntra Auth SDK (sidecar) / workload identity federationCSA ATF element 1 (Identity)GA 2026
Conditional access (delegated)Conditional Access for agentsCSA ATF element 4 (Segmentation)GA 2026
Conditional access (application-only)Conditional Access for agentsCSA ATF element 4 (Segmentation)GA 2026
Risk-based blockingIdentity Protection for agents (agent risk)CSA ATF element 2 (Behaviour monitoring)Preview (agent risk); service plans rolled out Jul-Aug 2026
Least privilege at tool callTool-scoped permissions; delegation constraintOWASP excessive agency; CSA recursive least privilegeConfiguration
Just-in-time credentialsPurpose-bound, auto-expiring credentialsCSA ATF (JIT provisioning, auto-expiry)Configuration
Audit loggingEntra sign-in and audit logs; unique Agent IDCSA ATF element 2; OWASPGA 2026
Data governanceMicrosoft Purview (DLP, sensitivity labels, DSPM for AI)CSA ATF element 3 (Data governance)GA 2026
Threat detection / responseMicrosoft Defender; investigation packageCSA ATF element 5 (Incident response)GA 2026
Revocation / kill-switchIdentity governance lifecycle; rapid-response CA policyCSA ATF element 5 (Incident response)GA 2026

Sources: Mi­crosoft Learn and Mi­crosoft an­nounce­ments, 2026; Cloud Se­cur­ity Al­li­ance Agen­tic Trust Frame­work,

Feb­ru­ary 2026; OWASP Top 10 for Agen­tic Ap­plic­a­tions, Decem­ber 2025.

Appendix B , Identity-governance checklist for agents

The fol­low­ing is a min­imum iden­tity con­trol set, drawn from Mi­crosoft Entra Agent ID and Agent 365 guid­ance (2026), the OWASP Top 10 for Agen­tic Ap­plic­a­tions (Decem­ber 2025), and the Cloud Se­cur­ity Al­li­ance Agen­tic Trust Frame­work (Feb­ru­ary 2026). Each block closes one step of the breach pipeline in chapter one.

Iden­tity and in­vent­ory

  • Every agent holds a first-class, cryp­to­graphic Entra Agent ID iden­tity; none au­then­tic­ates with a shared key or a human cre­den­tial.
  • Every agent is re­cor­ded in the Agent 365 re­gistry, in­clud­ing third-party and pre­vi­ously shadow agents.
  • Each agent is clas­si­fied as del­eg­ated-ac­cess or own-ac­cess; own-ac­cess is re­served for genu­inely autonom­ous agents.
  • Agent blue­prints carry baseline se­cur­ity policy so scope is con­sist­ent across in­stan­ti­ated iden­tit­ies.

Least priv­ilege

  • Tool per­mis­sions are scoped per agent to the spe­cific task; no agent holds broad tool ac­cess by de­fault.
  • Del­eg­a­tion in­her­it­ance is con­strained so a sub-agent can­not ex­ceed its par­ent's scope.
  • Cre­den­tials are just-in-time, pur­pose-bound, and auto-ex­pir­ing; no stand­ing secrets per­sist past the task.
  • Ac­cess is with­drawn on task com­ple­tion, not at in­def­in­ite fu­ture re­view.

Con­di­tional ac­cess and risk

  • A con­di­tional-ac­cess policy blocks every agent ex­cept those car­ry­ing an ap­proved cus­tom se­cur­ity at­trib­ute.
  • A con­di­tional-ac­cess policy blocks any agent rated high-risk by Iden­tity Pro­tec­tion.
  • Both policies were val­id­ated in re­port-only mode be­fore en­force­ment.

Audit and re­voc­a­tion

  • Every au­then­tic­a­tion, token ac­quis­i­tion, tool in­voc­a­tion, and policy de­cision is logged with a unique Agent iden­ti­fier.
  • Entra logs are in­teg­rated with Pur­view (data) and De­fender (threat) for uni­fied in­vest­ig­a­tion.
Zero Trust SecurityVoice of Volge20 •
  • The kill-switch has been tested against a live agent, con­firm­ing loss of ac­cess, token in­val­id­a­tion, and log­ging.
  • Agent off­board­ing is part of the iden­tity-gov­ernance li­fe­cycle, not an ex­cep­tional manual act.

Appendix E: the regional position, layer by layer, as at 8 November 2026

LayerRegional positionSource
Entra ID and Entra multifactor authenticationNamed as services that may store data globally, as exceptions to the tenant's chosen geography. Not in region anywhereMicrosoft Learn, Copilot Studio and Dynamics 365 data residency pages, updated 30 April and 7 October 2026
Agent 365, registry and agent tracesTenant geo home location only. Macro region and environment-level region selection do not applyMicrosoft Learn, Dynamics 365 and Power Platform data residency, updated 7 October 2026
Microsoft 365 core workloadsDurable in-country commitments in India, Qatar and the UAE. No commitment for Saudi Arabia, Kuwait, Oman or BahrainMicrosoft Learn, Compare Microsoft 365 Data Residency offerings, updated 27 October 2026
Model inferenceRouted to the closest datacentres in the region, and may call into other regions where capacity is available. Only the EU Data Boundary is a processing commitmentMicrosoft Learn, Data, Privacy and Security for Microsoft Copilot, updated 3 November 2026
Copilot Studio agent dataUAE North and two Indian regions listed. Qatar, Saudi Arabia, Kuwait, Oman and Bahrain not listed, so data is stored in the United StatesMicrosoft Learn, Data locations in Copilot Studio, updated 30 April 2026
Regulation of non-human identityNo instrument in any of the seven jurisdictions regulates machine identity as a categoryVerified negative from exhaustive text search of the primary instruments

Com­piled by Volge, 8 Novem­ber 2026. Product cap­ab­il­it­ies and data loc­a­tions change without no­tice; verify against cur­rent vendor doc­u­ment­a­tion be­fore re­ly­ing on this table. Not legal ad­vice.

Pre­pared by Volge · June 2026, re­vised Novem­ber 2026. This doc­u­ment is a point of view in­ten­ded to in­form se­cur­ity and iden­tity de­cision-mak­ing. It is not legal, se­cur­ity, or com­pli­ance ad­vice. Stat­ist­ics are at­trib­uted to their sources and were ac­cur­ate to the best of avail­able know­ledge as at Novem­ber 2026; read­ers are ad­vised that the agen­tic-AI and iden­tity-se­cur­ity field is mov­ing rap­idly and fig­ures should be re­con­firmed be­fore they are re­lied upon. Product cap­ab­il­it­ies, avail­ab­il­ity dates, and li­cens­ing are sub­ject to change by the vendor; con­fig­ur­a­tions should be val­id­ated against cur­rent vendor doc­u­ment­a­tion. Qual­i­fied se­cur­ity coun­sel should be en­gaged for de­cisions bear­ing on reg­u­lat­ory or con­trac­tual ob­lig­a­tions.

Zero Trust SecurityVoice of Volge21 •

Reference

Glossary
  • Agent identity blueprint. An object in Microsoft Entra ID that serves as the template and authentication foundation for one or more agent identities; it holds the credentials and acquires their tokens, and policy set on it covers every agent identity created from it.
  • Agent identity. An instantiated, first-class identity in Entra Agent ID that acts against resources, authenticating with tokens issued by its blueprint rather than with credentials of its own.
  • Agent's user account. An optional account, paired one to one with an agent identity, used where an agent experience requires a user object; it does not replace the agent identity, and it can hold neither passwords nor privileged administrative roles.
  • Agent registry. The inventory of every agent in the tenant, consolidated into Microsoft Agent 365 on 1 May 2026.
  • Agent 365. Microsoft's unified control plane for agent visibility, governance, and security; generally available 1 May 2026.
  • Conditional Access for agents. The Entra policy engine extended to evaluate every token an agent acquires against scope, approval attribute, and agent risk.
  • Delegated-access agent. An agent that acts on behalf of a signed-in user and acquires access in that user's context.
  • Entra Agent ID. Microsoft's identity and security framework assigning a first-class identity to each agent; generally available April 2026.
  • Excessive agency. The OWASP-named risk in which an agent holds more permissions, functionality, or autonomy than its task requires.
  • Just-in-time credential. A temporary, scoped, purpose-bound credential that expires automatically when the task completes.
  • Non-human identity (NHI). A machine, service, or agent credential; NHIs outnumber human users by 109 to 1 in the enterprises surveyed for the 2026 Identity Security Landscape.
  • Own-access agent. An agent that operates with its own identity, not tied to any user session, typically autonomous and event-triggered.
  • Zero Trust. A security posture in which no actor, human or agent, is trusted by default and trust is continuously verified.
Zero Trust SecurityVoice of Volge22 •
Sources

This paper draws on the fol­low­ing, ac­cessed June 2026. Fig­ures are at­trib­uted in-text to their primary source. Where a stat­istic ap­pears across mul­tiple sec­ond­ary out­lets, the most au­thor­it­at­ive primary at­tri­bu­tion is given.

  • Microsoft , Entra Agent ID (general availability April 2026), Agent 365 (general availability 1 May 2026, USD 15 per user per month standalone or inside Microsoft 365 E7), agent registry convergence, Conditional Access for agents, Identity Protection for agents, and Purview and Defender for agent governance; Microsoft Learn documentation, 2026.
  • Palo Alto Networks , 2026 Identity Security Landscape (survey of 2,930 cybersecurity leaders), on machine identities outnumbering human identities by 109 to one. The series was published under the CyberArk name until Palo Alto Networks completed its acquisition of CyberArk in February 2026.
  • Cloud Security Alliance , Agentic Trust Framework (2 February 2026), authored by Josh Woodruff of MassiveScale.AI; five core elements and a four-level maturity model.
  • Cloud Security Alliance , Securing Autonomous AI Agents (5 February 2026), a survey of 285 IT and security professionals commissioned and financed by Strata Identity and fielded online in September and October 2025; audit doubt, confidence in existing identity and access management, registry coverage, and traceability to a human.
  • OWASP Gen AI Security Project , Top 10 for Agentic Applications (9 December 2025), branded the 2026 edition and identifying its risks ASI01 to ASI10, with input from over 100 contributors; excessive agency, insecure tool execution, and related risks.
  • Entro Labs (cloud-native and DevOps identity ratios, first half of 2025), ManageEngine 2026 Identity Security Outlook (sector-level ratio maximum), and Enterprise Management Associates (September 2026, agents acting outside intended scope, commissioned by Cequence Security).
  • Microsoft 365 Message Center MC1395007 (17 June 2026): Entra Conditional Access for Agents and Entra ID Protection for Agents service plans added to Microsoft Agent 365 and Microsoft 365 E7, rollout July to August 2026, an entitlement notice rather than an availability notice. Microsoft Learn, Conditional Access for agents and Secure autonomous agents with Conditional Access, 2026, for the access patterns and their licensing.
  • Volge flagship , From copilot to autopilot: operationalising autonomous AI agents in the enterprise (June 2026), Voice of Volge series, for the board-level identity and governance position this paper extends.
Zero Trust SecurityVoice of Volge23 •

Volge and our expertise on agent identity

Volge builds gov­erned agent es­tates in­side the cli­ent's own Entra ten­ant, and the work is iden­tity en­gin­eer­ing rather than ad­vice. The prac­tice cov­ers the five lay­ers this paper sets out, from the Entra Agent ID ob­ject model and the Agent 365 re­gistry through con­di­tional-ac­cess design, the least-priv­ilege cre­den­tial li­fe­cycle, the audit in­teg­ra­tion with Pur­view and De­fender, and the tested kill-switch. The de­liv­er­able is an es­tate that can demon­strate con­trol of its agents to an aud­itor.

The en­gage­ment fol­lows the paper's own order, be­cause the order is the dis­cip­line. Iden­tity and in­vent­ory come first, built by re­con­cil­ing the plat­form view, the dir­ect­ory, the billing re­cord and the net­work egress against each other. Policy fol­lows, with both ca­non­ical policies val­id­ated in re­port-only and pro­moted on a dated plan, and with both policy sub­jects covered rather than one. Then audit, proven by re­con­struct­ing a week of one agent's ac­tions from logs alone. Then re­voc­a­tion, proven in a drill against a live agent in a con­trolled win­dow.

The re­gional work is handled ex­pli­citly rather than in­her­ited from a tem­plate. Volge is headquartered in the Gulf, and the layer-by-layer po­s­i­tion in chapter five is es­tab­lished in writ­ing be­fore the design is fixed, in­clud­ing the Agent 365 trace loc­a­tion, the Co­pi­lot Stu­dio data loc­a­tion and the Entra ex­cep­tion. Where no in­stru­ment defines a ma­chine iden­tity, the prac­tice writes the con­trol, maps it to the OWASP and Cloud Se­cur­ity Al­li­ance ref­er­ences an aud­itor re­cog­nises, and takes it through the cli­ent's own gov­ernance so that the de­cision is re­cor­ded rather than as­sumed.

Who does the work. A de­ploy­ment strategist holds the se­cur­ity out­come and the gov­ernance ca­dence with the cli­ent's chief in­form­a­tion se­cur­ity of­ficer. A trans­form­a­tion ar­chi­tect holds the con­trol design and the se­quence. Senior iden­tity and se­cur­ity en­gin­eers con­fig­ure in­side the cli­ent's ten­ant along­side the cli­ent's own team, and stay through the cutover from re­port-only to en­force­ment. The same people re­main from as­sess­ment to han­dover, which is why the con­trol tests in chapter three are drills the team has run rather than pro­ced­ures it has re­com­men­ded.

An en­gage­ment starts with a fixed-scope pos­ture as­sess­ment, usu­ally four to six weeks. It re­turns an in­vent­ory of the live agent es­tate, un­re­gistered and third-party agents in­cluded, each agent clas­si­fied as del­eg­ated-ac­cess or own-ac­cess, a place­ment on the ma­tur­ity model in chapter four, the li­cence en­ti­tle­ment po­s­i­tion, and a se­quenced route to the audit-ready level with the con­trol test that closes each gap. Volge is a Mi­crosoft part­ner and be­ne­fits com­mer­cially where a cli­ent ex­pands its Mi­crosoft and Entra es­tate, which a reader should weigh when read­ing the re­com­mend­a­tions here. The con­ver­sa­tion be­gins with whatever in­vent­ory ex­ists today, how­ever in­com­plete.

Zero Trust SecurityVoice of Volge24 •
Version · VVM-ZTS-003 · 2026
volgeEnable Powerful
Transformations
About Volge

Volge is an engineering firm providing enterprise technology transformation services to clients across the Middle East, India, APAC, Europe, and North America.

Delivery stands on two pillars: Forward Deployed Engineering, with senior specialists embedded in the client’s own team, and Intelligent Command Centre Operations, run from our Global Delivery Centres, which keeps what they build running. Senior engineers stay with an engagement from the first diagnosis through to production.

Our services span Strategy, Data, Platforms, AI and Automation, Cyber Trust, Technology Operations, and Applied Functional Engineering. Engagements begin with a diagnosis. The firm is built on working autonomy: engineers decide, innovate, co-create, and own their work. Motivated engineers build the systems that sustain our clients’ growth.

Terms of use

This edition may be shared with organisations evaluating a Volge engagement. It may not be altered or republished without Volge’s permission, and figures carry their stated basis and date. Volge, Intelligent Command Centre, and Intelligent Command Centre Operations are marks of Volge; other names are the trademarks of their respective owners.