Safety#AI safety

Proofpoint details BluePhishing agent consent phishing proof of concept

Proofpoint's Nov. 9 report describes BluePhishing, a consent-phishing proof of concept that could create Entra agent identities. No live attacks were observed.

An abstract network texture of black, white and green points of light

Proofpoint published its BluePhishing research on November 9, 2026. It describes a proof of concept for phishing consent to a third-party blueprint in Microsoft Entra Agent ID: an attacker persuades an employee to approve the blueprint, then uses the granted capability to create linked agent identities in the victim tenant. Proofpoint said it had seen no activity in the wild at publication, so the report documents a tested attack path, not a confirmed breach.

Key points

  • BluePhishing places the consent decision inside what can look like an ordinary third-party agent onboarding flow.
  • In Proofpoint’s test, an approved blueprint could create agent identities linked to it in the consenting organization’s tenant.
  • Proofpoint called the work a proof of concept and reported no in-the-wild attacks as of November 9; TechRadar cited no victim or independent forensic evidence.

Background

Microsoft Entra Agent ID gives organizations identity and authorization controls for AI agents. Microsoft describes an agent identity as a special service principal created from a reusable agent identity blueprint. The identity has no credentials of its own; the blueprint holds credentials and obtains tokens on its behalf. A blueprint can also define shared metadata and permissions, allowing administrators to manage access controls for agents of the same kind.

That model explains how an organization can manage agents through its identity system. It does not establish that every agent is trustworthy. A provider can configure a blueprint for multiple tenants, so it can be added to other organizations and create tenant-local identities after consent. The consent page and its Microsoft flow may be genuine. The security decision is which publisher receives approval, what permissions it receives and how those permissions can later be used.

Organizations already contend with OAuth consent phishing, where employees are tricked into authorizing a third-party application that then accesses data within its granted scopes. Our report on fake event invitations harvesting work accounts covered a related trust mechanism, though that campaign used device-code and reverse-proxy techniques to obtain account tokens. BluePhishing concerns blueprint consent and derived agent identities; the paths differ, so the risks should not be conflated.

Proofpoint’s scenario starts with an agent blueprint prepared by an attacker. The attacker configures it for acceptance by multiple Entra tenants, then uses social engineering to get a target employee to open the authorization page and approve the request. If consent succeeds, the target tenant creates a local object linked to the blueprint and grants the capability needed to create agent identities. The blueprint can then create an identity in that tenant; its access depends on the permissions granted earlier.

Consent is the critical boundary in this chain. The attacker is using legitimate authorization and identity-creation features, not an Entra platform vulnerability. Proofpoint says a blueprint can carry or pass on granted permissions. Investigators therefore need to review more than the direct permissions on an agent identity: they need the source blueprint, consent record and inheritance relationship. The report also says some inherited permissions may not be obvious in common management views, which can leave investigators short of context.

That does not mean every approval of a third-party agent leads to compromise. Outcomes depend on the scopes approved, tenant consent policy, administrator controls and the resource’s access rules. Broad application permissions generally require administrator consent; user consent is constrained by organizational policy. An attacker would also need control of the blueprint’s credentials and satisfy the later identity-creation and token-use requirements. Proofpoint demonstrates how those conditions can fit together, not a real intrusion.

Entra’s model

Microsoft documents an agent identity blueprint as a reusable template for creating agent identities. Each identity is a special service principal in a tenant, with its own identifier but no credentials of its own. The blueprint uses its credentials to request tokens for the identity. An agent identity can authenticate to APIs and receive tokens addressed to it; an interactive agent can also request a user token when that user has authorized the relevant scopes.

The division between blueprint and identity explains the research’s potential reach. Microsoft says an agent identity can receive tokens only in the tenant where it was created, while a blueprint can be configured for multiple tenants and create local identities in each. The design lets administrators manage a class of agents through its blueprint. It also makes cross-tenant consent an entry point that needs governance. Proofpoint discusses a malicious use of that capability; Microsoft’s documentation describes the product design and does not endorse the report’s conclusions.

For administrators, the blueprint identifier, publisher, granted Graph permissions, derived identities and sponsor together provide investigative context. Microsoft says a sponsor records the user or group accountable for an agent and can help identify whom to contact during an incident. Correlating these objects is more reliable than searching by display name alone: names are readable and can change, while stable identifiers and tenant scope help comparison.

Independent checks

TechRadar reported on Proofpoint’s research on November 10. It relayed the concern that an approved third-party blueprint could create agent identities and repeated that Proofpoint had no evidence of live attacks. The article confirms that security media covered the disclosure, but it relies on the same Proofpoint material and does not supply a second team’s reproduction or a victim case.

Swiss security commentary site CTI Pilot reviewed the report’s evidence boundary. It says the permission behavior and technical conclusions still rest on Proofpoint’s own testing, with no independent confirmation and no public Microsoft response about a fix; it also notes that Proofpoint notified Microsoft before publication. There is a second author assessing the evidence, but no second vendor or research team has validated the attack chain. Microsoft’s official documentation explains Agent ID’s design and does not mention BluePhishing or respond to the research.

The material available so far also identifies no victims, stolen data or telemetry showing attackers using the technique. Calling a feasible proof of concept an active campaign would overstate the evidence. A reproducible independent analysis, a Microsoft notice or tenant audit data could change that assessment; the current sources do not support an estimate of how often the risk occurs.

Defensive implications

For Entra administrators and identity teams, the immediate work is to review consent boundaries for third-party agents and applications. Limit the permissions users can approve by default, require administrator review for high-impact scopes and disable user consent where the organization does not need it. When a blueprint is introduced, verify its publisher, requested resource permissions, business owner and deployment scope. Consent records should map to the specific blueprint and the agent identities it later creates.

Proofpoint recommends monitoring audit events tied to agent blueprints, consent and agent identity creation, while applying least privilege and Conditional Access policies to agents. Incident responders should correlate blueprint IDs, tenant objects, grants, sponsors and related token activity on one timeline. If the same blueprint appears across tenants, its stable identifier can help defenders spot it; that identifier is a lead, and still needs to be assessed alongside the publisher and permission changes.

The evidence has a clear limit: Proofpoint had observed no real-world activity at publication, and CTI Pilot found neither an independent reproduction nor a public Microsoft response. Organizations can use the report to review consent policy and log coverage, while treating it as a warning about a possible path rather than proof of compromise. For employees, the practical rule is to treat an authorization screen as a security decision: check the agent publisher, business purpose and requested permissions, and send unclear requests to an administrator.