>
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.
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.
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.
how a single identity gap becomes a data breach
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 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.
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
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).
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.
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).
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 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).
| Estate type | Non-human to human identity ratio | Source basis |
|---|---|---|
| Enterprise average | 109 : 1 | Palo Alto Networks, 2026 |
| Of which, agent identities | 79 : 1 | Palo Alto Networks, 2026 |
| Cloud-native / DevOps | 144 : 1 | Entro Labs, first half of 2025 |
| Sector maximum | Up to 500 : 1 | ManageEngine, 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 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).
| Capability | Share of organisations | Source |
|---|---|---|
| Doubt they could pass an audit of agent behaviour or access | 84% | Cloud Security Alliance, February 2026 |
| Highly confident existing IAM can manage agent identities | 18% | Cloud Security Alliance, February 2026 |
| Maintain a real-time registry of agents | 21% | Cloud Security Alliance, February 2026 |
| Can trace agent actions to a human across all environments | 28% | 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 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.
identity as the control plane
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.
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
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.
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).
| Object | What it is | Function |
|---|---|---|
| Agent identity blueprint | An object in Microsoft Entra ID serving as the template and authentication foundation for one or more agent identities | Holds 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 principal | The object recording a blueprint's presence in the tenant | Lets 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 identity | The primary identity an agent uses to authenticate | Holds no credentials of its own and authenticates with tokens issued by its blueprint; the actor in agentic tasks |
| Agent's user account | An optional account pairing one to one with an agent identity | Used 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).
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.
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).
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.
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).
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.
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.
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.
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.
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.
scope, conditional access, audit, and revocation
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 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).
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.
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.
| Dimension | Prevailing practice | Least-privilege practice |
|---|---|---|
| Credential | Standing secret, no expiry | Just-in-time, auto-expiring, purpose-bound |
| Scope unit | The agent, set once at deployment | Each tool call, evaluated per task |
| Delegation | Inherited broadly down the chain | Constrained at every step; sub-agent ≤ parent |
| Permission review | Never revisited | Continuous; access withdrawn on task completion |
| Identity basis | Shared key or human credential | First-class agent identity |
Sources: OWASP Top 10 for Agentic Applications, December 2025; Cloud Security Alliance Agentic Trust Framework, February 2026.
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
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.
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).
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.
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 question | Where answered | Mechanism |
|---|---|---|
| Which agent acted? | Sign-in / audit logs | Unique Agent ID per action |
| On whose behalf? | Sign-in logs | Delegated vs own-access flow record |
| Under what policy? | Conditional Access logs | Policy evaluation result, including report-only |
| What data was touched? | Purview audit / DSPM for AI | DLP, sensitivity labels, classification |
| Was it anomalous? | Identity Protection | Agent risk score; Defender verdict |
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.
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.
| Control | Where it is configured | Test | Signature of failure |
|---|---|---|---|
| Dedicated agent identity | Entra Agent ID blueprint and agent identity | Trace one action from tool call to sign-in record and confirm a unique agent identifier | The action resolves to a shared service principal or to a human account |
| Registry completeness | Agent registry in Microsoft Agent 365 | Reconcile the registry against directory token activity, billing and network egress | An agent with recent activity appears in only one of the four sources |
| Tool-level least privilege | Tool permissions per agent | Invoke a tool outside the agent's task scope | The call succeeds, or fails without a log entry |
| Delegation constraint | Sub-agent permission inheritance | Reach a parent-forbidden resource through a sub-agent | The chain succeeds, and effective reach exceeds the documented grant |
| Credential lifetime | Just-in-time, purpose-bound credentials | Complete a task, then re-attempt authentication with the issued credential | The credential still authenticates after the task has ended |
| Approval gate | Conditional Access on a custom security attribute, with a second policy for agent user accounts | Create an agent without the approved attribute value and attempt resource access | Access succeeds, usually because only one of the two subjects was covered |
| Risk-based block | Conditional Access on the agent risk condition | Raise a test risk signal and confirm the block, in report-only first | No evaluation record, or an evaluation that allows |
| Audit sufficiency | Entra sign-in and audit logs, with Purview and Defender integration | Reconstruct a week of one agent's actions from logs alone | A gap at what was done, on whose behalf, under what policy, or why |
| Revocation | Identity governance lifecycle and a rapid-response policy | Fire the kill-switch against a live agent in a controlled window | Access 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.
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.
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.
reference architecture, rollout, and maturity
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 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
(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 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).
| Capability | Licence required |
|---|---|
| Agent identity platform (create and manage agent identities) | Available to all Microsoft Entra customers |
| Operate agents across Microsoft 365 and enterprise workflows | Microsoft Agent 365 licence per user |
| Conditional Access for agents | Microsoft 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 agents | Microsoft Agent 365 or Microsoft 365 E7 (Entra ID Protection for Agents service plan) |
| Identity Governance for agents | Microsoft Entra ID P1 |
| Network controls for agents | Microsoft Entra Internet Access (Entra Suite or separate) |
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.
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
never reversed an enforcement has probably never enforced at scale.
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.
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.
A stage is complete when it can produce evidence, not when the configuration is saved. Four measures and four artefacts settle it.
| Stage | Measure | Artefact |
|---|---|---|
| One, identity and inventory | Share of agents holding a dedicated identity, and share present in the registry, against the reconciled population | A reconciliation record naming every difference between the four discovery sources, and its resolution |
| Two, policy and least privilege | Share of agents scoped at the tool, share on just-in-time credentials, and the count of policies promoted out of report-only | Both canonical policies in enforcement, each with the dated report-only impact review behind it |
| Three, audit and integration | Sampled share of agent actions reconstructable from logs alone | One completed investigation package assembled from Entra, Purview and Defender for a single agent action |
| Four, revocation and operations | Time from decision to confirmed loss of access, measured in a drill | A 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.
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.
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 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
chapter one: Level 0 is the breach pipeline running unbroken; each higher level breaks one more step.
| Level | Identity state | Access control | Audit and revocation | Typical exposure |
|---|---|---|---|---|
| 0 , Invisible | Agents use shared keys or human credentials; no inventory | None scoped to agents | No agent-specific logging; no kill-switch | Rogue exposure; cannot pass audit |
| 1 , Identified | Every agent has a first-class Entra Agent ID identity; registered in Agent 365 | Inherited or broad | Agent actions logged with unique ID | Visible but over-permissioned |
| 2 , Scoped | First-class identity; delegated vs own-access classified | Conditional Access enforced; least privilege at tool and credential | Full sign-in and audit logging; revocation defined | Controlled; audit-ready |
| 3 , Verified | Identity plus continuous risk evaluation | Risk-adaptive Conditional Access; JIT credentials estate-wide | Purview and Defender integrated; kill-switch tested | Low; continuously verified |
| 4 , Governed estate | Estate-wide identity spanning Microsoft and third-party agents | Policy spans vendors; delegation constrained end-to-end | Unified investigation; routine offboarding | Minimal; 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 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.
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.
| Jurisdiction | Microsoft 365 Local Region Geography | Agent 365 trace location | Entra ID and multifactor authentication data |
|---|---|---|---|
| India, Qatar, United Arab Emirates | Yes, with durable commitments on the six core workloads | Tenant home geography, in country | May be stored globally |
| Saudi Arabia, Kuwait, Oman, Bahrain | No, and not eligible for the Advanced Data Residency add-on | Tenant home geography, with no commitment to a location | May 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.
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).
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.
Saudi Arabia's identity assets, Nafath and Tawakkalna, are human-identity services and should not be cited as a
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.
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 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.
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.
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 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 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
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.
The following maps each identity control to its Microsoft capability and its framework reference, as at November 2026.
| Control | Microsoft capability | Framework reference | Availability |
|---|---|---|---|
| First-class agent identity | Entra Agent ID (agent identity blueprint, agent identity, optional agent's user account) | CSA ATF element 1 (Identity); OWASP excessive agency | GA April 2026 |
| Agent inventory / registry | Agent registry in Microsoft Agent 365 | CSA ATF element 1 (Identity) | GA 1 May 2026 |
| Third-party agent onboarding | Entra Auth SDK (sidecar) / workload identity federation | CSA ATF element 1 (Identity) | GA 2026 |
| Conditional access (delegated) | Conditional Access for agents | CSA ATF element 4 (Segmentation) | GA 2026 |
| Conditional access (application-only) | Conditional Access for agents | CSA ATF element 4 (Segmentation) | GA 2026 |
| Risk-based blocking | Identity 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 call | Tool-scoped permissions; delegation constraint | OWASP excessive agency; CSA recursive least privilege | Configuration |
| Just-in-time credentials | Purpose-bound, auto-expiring credentials | CSA ATF (JIT provisioning, auto-expiry) | Configuration |
| Audit logging | Entra sign-in and audit logs; unique Agent ID | CSA ATF element 2; OWASP | GA 2026 |
| Data governance | Microsoft Purview (DLP, sensitivity labels, DSPM for AI) | CSA ATF element 3 (Data governance) | GA 2026 |
| Threat detection / response | Microsoft Defender; investigation package | CSA ATF element 5 (Incident response) | GA 2026 |
| Revocation / kill-switch | Identity governance lifecycle; rapid-response CA policy | CSA 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.
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
Least privilege
Conditional access and risk
Audit and revocation
| Layer | Regional position | Source |
|---|---|---|
| Entra ID and Entra multifactor authentication | Named as services that may store data globally, as exceptions to the tenant's chosen geography. Not in region anywhere | Microsoft Learn, Copilot Studio and Dynamics 365 data residency pages, updated 30 April and 7 October 2026 |
| Agent 365, registry and agent traces | Tenant geo home location only. Macro region and environment-level region selection do not apply | Microsoft Learn, Dynamics 365 and Power Platform data residency, updated 7 October 2026 |
| Microsoft 365 core workloads | Durable in-country commitments in India, Qatar and the UAE. No commitment for Saudi Arabia, Kuwait, Oman or Bahrain | Microsoft Learn, Compare Microsoft 365 Data Residency offerings, updated 27 October 2026 |
| Model inference | Routed 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 commitment | Microsoft Learn, Data, Privacy and Security for Microsoft Copilot, updated 3 November 2026 |
| Copilot Studio agent data | UAE North and two Indian regions listed. Qatar, Saudi Arabia, Kuwait, Oman and Bahrain not listed, so data is stored in the United States | Microsoft Learn, Data locations in Copilot Studio, updated 30 April 2026 |
| Regulation of non-human identity | No instrument in any of the seven jurisdictions regulates machine identity as a category | Verified 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.
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.
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.
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.
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.