BluePhishing Attack: How Microsoft Entra Agent Identities Could Be Exploited

BluePhishing Attack: How Microsoft Entra Agent Identities Could Be Exploited

BluePhishing Exposes an Emerging Identity Security Gap in Microsoft’s AI Agent Framework

A Proofpoint proof of concept demonstrates how consent to a third-party agent blueprint could provide an attacker with a route into an enterprise Microsoft Entra tenant. The research highlights risks around inherited permissions, agent provisioning and identity governance—but recent Microsoft changes make the attack’s operational conditions more nuanced than the initial disclosure suggests.

AI agents are acquiring identities, credentials and access rights inside enterprise environments. That gives organisations a way to govern autonomous software more systematically, but it also creates another identity surface that security teams must protect.

A new proof of concept from Proofpoint illustrates how attackers could exploit that surface without stealing an employee’s password or presenting a counterfeit Microsoft login page.

The technique, named BluePhishing, targets Microsoft Entra Agent ID and its agent identity blueprint model. It adapts a familiar attack pattern—OAuth consent phishing—to a newer identity architecture designed to provision and manage AI agents.

In its 9 November 2026 research disclosure, Proofpoint describes how a malicious third-party blueprint could persuade a user or administrator to grant consent inside a victim organisation’s Microsoft Entra tenant. The attacker could then use the resulting blueprint principal to create agent identities and obtain tokens carrying permissions associated with those identities.

The important distinction is that this is a demonstrated research technique, not a confirmed campaign. Proofpoint explicitly states that it has not identified BluePhishing activity in the wild. The research establishes a potential attack path through documented platform capabilities; it does not establish that attackers are currently exploiting it at scale.

Nevertheless, the disclosure deserves attention because it brings together three enterprise security challenges: consent governance, autonomous identity provisioning and permission inheritance. It also arrives as Microsoft changes the default controls governing who can create agent blueprints and how those blueprints obtain authority to create identities.

How Microsoft Entra Agent ID creates a new identity layer

Traditional enterprise identity management revolves around users, applications, service principals and managed identities. AI agents introduce another category: identities associated with autonomous or semi-autonomous software entities.

Microsoft Entra Agent ID extends the Entra identity framework to these agents. The objective is to give organisations a way to identify them, assign permissions, establish accountability and monitor their activity.

An agent identity blueprint acts as a template. It defines configuration and authentication properties that can be shared by the agent identities created from it. The architecture distinguishes three important objects:

  • Agent identity blueprint: The template that defines how agent identities are configured and authenticated.
  • Blueprint principal: The tenant-local representation of a multi-instance blueprint. It allows the blueprint to operate within the organisation’s Entra environment.
  • Agent identity: The individual identity created from the blueprint and used by an agent when accessing resources.

The model has similarities to the traditional application registration and service principal relationship. An application registration defines an application, while its service principal represents that application in a particular tenant. In the agent identity model, the blueprint and blueprint principal perform comparable roles.

The distinction matters because a third-party blueprint can be created in one organisation and configured for use in another. A customer organisation that consents to the blueprint obtains a corresponding principal in its own tenant.

Microsoft supports single-instance and multi-instance blueprint modes. A single-instance blueprint represents one agent identity in a tenant. A multi-instance blueprint can provision multiple identities, each with its own identity and permissions, while sharing authentication settings configured at the blueprint level.

That multi-instance capability is central to BluePhishing. It allows a blueprint to act as a parent for multiple identities rather than limiting the attacker’s opportunity to a single application-like identity.

There is, however, an important product-maturity qualification. Microsoft’s documentation identifies agent identity blueprint instance modes as a preview capability, without a service-level agreement, and says they are not recommended for production workloads. This makes the security implications particularly relevant to organisations evaluating or experimenting with the architecture, but it also means administrators should distinguish preview functionality from a fully mature production control plane.

Microsoft’s documentation on agent identity blueprints describes the object relationships and the differences between the two instance modes.

BluePhishing: Turning legitimate consent into an entry point

BluePhishing builds on OAuth consent phishing, a technique in which an attacker persuades a user to authorise a malicious application to access organisational resources.

The conventional attack does not necessarily require the victim to disclose a password. Instead, the victim grants an application permissions through a legitimate identity-provider consent flow. The application then uses the granted authorisation to access resources.

BluePhishing applies this principle to third-party agent blueprints.

According to Proofpoint, an attacker first creates an agent identity blueprint in an Entra tenant under their control. The attacker configures the blueprint for use across multiple organisations and adds a client credential. The blueprint is also configured with permission declarations and inheritable permissions.

The attacker then persuades someone in the target organisation to approve the blueprint through a consent link.

The sequence is significant:

  1. The attacker prepares a third-party agent blueprint in an attacker-controlled tenant.
  2. The blueprint is configured to support cross-tenant use and to request specified permissions.
  3. The attacker sends a consent link to a target user or administrator.
  4. The target grants consent, resulting in a blueprint principal being established in the victim tenant.
  5. The attacker attempts to authenticate as the blueprint in the victim tenant and create agent identities.
  6. Those identities can obtain tokens carrying permissions available through the blueprint’s grants and inheritance configuration.

The consent page is a genuine Microsoft authorisation experience. The victim is not necessarily being redirected to a counterfeit login page. The deception concerns the identity and trustworthiness of the agent being authorised, not the authenticity of the identity provider.

That changes the nature of the social-engineering problem. Users may be familiar with warnings about suspicious applications, but an agent blueprint can look like a legitimate component of an organisation’s growing AI ecosystem. An employee may see an agent as a productivity tool rather than as an application-level identity capable of receiving access to corporate resources.

Proofpoint says an attacker needs the blueprint application ID and client secret, together with the victim tenant ID and a sponsor user identifier. The tenant ID and user object identifier are not equivalent to authentication secrets. Their availability does not, by itself, compromise an organisation. They become useful when combined with a consent opportunity and the attacker’s ability to authenticate using the blueprint’s credentials.

The research also describes the use of Microsoft Graph to obtain tokens and create tenant-local agent identities. This is not an attack on Microsoft’s login infrastructure. It is an abuse scenario involving legitimate identity and authorisation mechanisms.

Proofpoint’s technical disclosure provides the proof-of-concept sequence and its prerequisites.

Why permission inheritance changes the risk

The most consequential part of the research is not simply that an attacker might register or introduce a malicious agent. It is the possibility that permissions granted to a blueprint can flow to identities created from it.

Microsoft designed inheritable permissions to simplify agent deployment. In an enterprise environment, a blueprint may be used to create multiple agent identities across teams, business units or workloads. Requiring administrators to approve the same permissions independently for every identity would make deployment cumbersome.

Inheritance addresses that problem. An administrator can approve permissions at the blueprint level, and eligible child identities can receive those permissions when they obtain access tokens.

Two configurations are important:

Required resource access (RRA) declares the APIs and permissions an agent needs. It provides administrators with a view of the permissions requested during static consent. RRA is a declaration, not a grant; authorisation still requires consent.

Inheritable permissions identifies the resource applications whose granted permissions can flow to child agent identities.

Microsoft’s documentation specifies that inheritance requires both conditions: the relevant scopes or roles must be configured as inheritable, and the permission must be granted through the applicable static or dynamic consent process.

The distinction between declared, granted and effective permissions is critical. A permission listed in a blueprint configuration is not automatically an active permission. It becomes effective only when the necessary grant exists and the inheritance conditions are met.

At token issuance, the platform can combine inherited and directly granted permissions. Delegated scopes appear in the scp claim, while application roles appear in the roles claim.

This creates an important monitoring challenge. Microsoft’s conceptual documentation says inherited permissions are not displayed as direct permissions on individual agent identities in the Entra admin centre or through Microsoft Graph. They are observable in the token contents at runtime.

That does not mean the permissions are entirely invisible to administrators. The blueprint principal and its grants remain important governance objects. Administrators can inspect blueprint permissions, review linked identities and examine sign-in and audit activity. The issue is that inspecting an individual child identity may not reveal its complete effective permission set as a direct grant.

There is also a documentation nuance worth noting. Microsoft’s conceptual page describes inheritance for both delegated scopes and application roles. Its administrative guidance discusses delegated-scope inheritance in more specific terms, while the public Graph beta reference does not fully document every application-role inheritance property. Organisations should therefore validate the precise inheritance behaviour supported by their tenant and API version rather than assume that every documented permission pattern is equally available through every management interface.

The practical security consequence remains clear: a review process that examines only direct permissions on an agent identity can produce an incomplete picture of what that identity can do at runtime.

The 250-identity claim and the potential blast radius

Proofpoint states that a single blueprint can be associated with up to 250 agent identities. This is a figure reported by the researchers; it should not be interpreted as an independently established universal production limit.

The number matters because it illustrates the potential scale of a blueprint-based attack.

A conventional application consent incident may create a malicious service principal in a victim tenant. BluePhishing introduces the possibility of using a consented blueprint as the parent of multiple agent identities.

Each child identity can represent a separate agent context. Depending on the configuration and permissions granted, those identities could be used for different tasks or access patterns.

That does not mean an attacker automatically receives 250 fully privileged identities after one consent action. The identities still depend on the platform’s provisioning rules, the blueprint’s authority, the available credentials, the sponsor requirements and the permissions actually granted.

Nor does the number of identities alone determine the impact. A single identity with highly privileged access may be more damaging than hundreds of identities with narrow permissions.

The more useful risk measure is the combination of:

  • The number of identities a blueprint can provision.
  • The permissions available to those identities.
  • The sensitivity of the resources they can access.
  • The ability to obtain and use tokens without interactive user involvement.
  • The organisation’s capacity to detect, investigate and revoke the relevant identities and grants.

For defenders, the blueprint becomes a critical point of control. If it is malicious or compromised, the organisation must assess not only the principal itself but also every identity created from it and every resource permission those identities can exercise.

Microsoft’s November changes alter the attack conditions

Two Microsoft Message Center announcements are especially relevant to the BluePhishing disclosure. They address separate aspects of agent identity governance and should be considered alongside Proofpoint’s proof of concept.

The first, MC1506733, concerns who can create agent blueprints. Microsoft announced that, beginning 4 November 2026, default users without built-in or custom Entra roles would be able to create agent blueprints unless administrators disabled the setting.

The relevant control is Users can create agent blueprints, located under Entra ID > Users > User settings. Administrators can set it to No to restrict default users from creating new blueprints. Supported administrators and users assigned the Agent ID Developer role can still create them.

The setting is separate from the familiar Users can register applications control. Organisations should review both rather than assume that restricting application registration also restricts agent blueprint creation.

Microsoft’s second announcement, MC1506728, concerns AgentIdentity.CreateAsManager, the Microsoft Graph application permission that allows a blueprint to create and manage agent identities.

This update is particularly important when assessing the original BluePhishing walkthrough.

Proofpoint’s research describes the blueprint principal receiving this permission implicitly, allowing it to create child identities without a separate administrator-granted app-role assignment. Microsoft subsequently announced that it would stop automatically granting this permission to agent blueprint principals, with rollout beginning in early November and expected to complete by late November 2026.

Under the revised approach, an administrator must explicitly grant the application permission to agents that require app-only authority to create identities. Microsoft says this makes the permission visible, auditable and revocable. Requests that rely only on the former automatic grant can fail with a 403 authorisation error after enforcement reaches the tenant.

This is a material change, not a minor administrative detail. The BluePhishing research was published on 9 November, while the Microsoft rollout was already under way. The original proof-of-concept sequence must therefore be evaluated against the controls active in the specific target tenant. A permission path demonstrated under the earlier behaviour should not be assumed to work unchanged in every tenant after the rollout.

Microsoft has also clarified in a Microsoft Q&A response that Microsoft-managed platforms such as Copilot and Foundry will continue to operate as required without the same explicit-consent requirement, while agents created in an organisation, another organisation or through third-party platforms will need explicit consent.

The distinction is important for enterprises. It is not a blanket prohibition on agent identity creation. It is a change in how authority is granted to create identities, with third-party and externally created agents facing a more explicit administrative approval requirement.

The rollout is expected to continue through late November. Administrators should verify the current behaviour in their own tenants rather than assume that every environment has already received the change.

The two announcements address different risks: MC1506733 governs who can create blueprints; MC1506728 governs the authority a blueprint principal needs to create identities. Neither should be treated as a substitute for reviewing third-party consent and permission grants.

Where defenders should focus: consent, identity inventory and runtime visibility

BluePhishing reinforces a broader identity-security principle: organisations must govern the authority being granted, not merely the identity requesting it.

A useful response begins with the consent process.

Restrict third-party consent. Organisations should review user-consent settings and prevent employees from approving untrusted applications or agent blueprints. Where user consent is allowed, permission classifications should limit which permissions users can approve. High-impact application permissions should remain subject to administrator review.

Control blueprint creation. Review the default-user blueprint creation setting and decide whether ordinary users need this capability. If agent development is centralised, restrict creation to designated administrators and developers. Maintain a separate review of application registration permissions.

Review every third-party blueprint. Verify the publisher, business purpose, owner, sponsor, requested permissions and intended deployment scope before granting consent. A familiar product name or an authentic Microsoft consent page is not sufficient evidence that a blueprint is trustworthy.

Inspect inherited permissions. Do not rely solely on the direct permission list shown for a child agent identity. Review the parent blueprint, its granted permissions and its inheritable-permission configuration. Where necessary, inspect access tokens in a controlled diagnostic process to understand the effective scp and roles claims.

Apply least privilege. A blueprint should inherit only the permissions required by its agents. Broad Graph permissions can create substantial exposure, particularly when identities can operate autonomously. Separate agent workloads and permission sets where their business purposes or sensitivity differ.

Monitor identity creation and sign-ins. Microsoft documents agent-related activity in Entra audit and sign-in logs. Administrators can inspect blueprint activity, agent identity creation events and sign-in records, including the agentSignIn resource type. Monitoring should correlate the creation of a blueprint principal with subsequent child identity creation and unexpected authentication activity.

Establish a response procedure. If a suspicious blueprint is identified, investigators should identify its linked agent identities, review their permissions and sign-ins, and determine whether tokens or credentials may have been used. Disabling a blueprint principal can prevent identities created from it from authenticating, according to Microsoft’s management documentation. Response teams should also assess related grants and credentials rather than treating the removal of one identity as complete remediation.

Microsoft’s documentation also identifies a specific Conditional Access limitation: Conditional Access enforcement does not apply when an agent identity blueprint acquires a token to create agent identities. Policies can still apply when an agent identity requests a token for a resource. This distinction matters because a control that protects the downstream agent sign-in does not necessarily govern the blueprint’s identity-provisioning operation.

The implication is that organisations need layered controls. Conditional Access, identity risk monitoring and audit logs remain relevant, but they should not be treated as complete protection for every stage of blueprint provisioning.

Microsoft’s agent identity management guidance and inheritable permissions documentation provide the administrative and technical details.

BluePhishing Attack: How Microsoft Entra Agent Identities Could Be Exploited
Photo by Jefferson Santos on Unsplash

The larger lesson: AI agent governance is identity governance

The BluePhishing research is not evidence that AI agents are inherently insecure. It demonstrates that a new identity framework can inherit familiar weaknesses when organisations extend trust without fully understanding the authority being granted.

The underlying problem is not unique to artificial intelligence. OAuth consent abuse has existed for years. What changes with agent identities is the provisioning model, the possibility of multiple child identities, the use of inherited permissions and the expectation that autonomous software will operate without continuous human interaction.

That combination can make malicious activity harder to distinguish from legitimate automation. An agent identity may authenticate and access resources as designed, even when the blueprint that created it was introduced through deception.

For security leaders, this shifts the question from whether an organisation has approved AI tools to whether it can account for every agent identity those tools create.

That requires a complete inventory of blueprints and child identities, clear ownership and sponsorship, a documented approval process for third-party agents, visibility into inherited permissions and monitoring of identity creation and token activity.

It also requires treating agent identity creation as a privileged operation. The November changes announced by Microsoft are a step towards making that authority more explicit, but they do not remove the need for organisations to review existing blueprints, permissions and consent policies.

Proofpoint has presented BluePhishing as a proof of concept and says it has not observed the technique in real-world attacks. That qualification should remain central to any assessment of the threat. The research is best understood as an early warning about how attackers could exploit the trust relationships being established around enterprise AI agents.

The next phase of enterprise AI adoption will depend not only on what agents can do, but on whether organisations can reliably establish who created them, who authorised them, what permissions they inherit and what they actually access.

An agent that has a legitimate identity is not automatically a trustworthy agent. The blueprint, the consent decision and the permissions behind that identity matter just as much.