Researchers show how malicious AI agents can enter corporate systems
Reported by TechRadar and WPRI 122 sources ↓
Proofpoint researchers found that attackers can use Microsoft Entra agent blueprints and social engineering to get malicious AI agents approved inside organizations.
Key points
- A victim who approves a malicious blueprint can add an attacker-controlled AI agent to their organization
- The approval link and Microsoft authorization page can be genuine
- TechRadar reports one blueprint can create up to 250 agent identities
Sources
Details
TechRadar reports that Proofpoint researchers described an attack in which a threat actor creates an agent blueprint in a Microsoft Entra environment they control, then persuades someone at a target organization to approve it. The link leads to a genuine Microsoft authorization page, so the approval request may not look like a conventional fake-login phishing attempt.
After approval, the agent can receive permission to create additional identities tied to the blueprint. TechRadar says a blueprint can support up to 250 agent identities; attackers can obtain access tokens for those identities and use their permissions to reach corporate resources.
The attack may require information about the target organization, including its tenant ID and details about a user to approach. Proofpoint advises restricting consent for third-party agents, limiting permissions, and monitoring agent and blueprint activity in audit logs.
Written from coverage as of
Images
- TechRadar
- WPRI 12
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh