Microsoft Adds 60 Agent Types to Entra ID for Insider Risk Control
The software giant warns that AI agents act like unpredictable insiders, requiring new identity and permission controls to prevent data exposure.
Key points
- Microsoft treats AI agents as insider threats due to their high-speed access to internal data and tools.
- The company is expanding Microsoft Entra Agent ID support to cover 60 distinct agent types.
- New controls separate the AI model’s intelligence from its authority to execute actions on company systems.
Microsoft advises organizations to treat artificial intelligence agents as potential insider security risks rather than assuming they are inherently malicious. The company published this warning on October 19, 2026, noting that agents can access sensitive corporate data and execute tasks at speeds that outpace human oversight. This guidance comes as businesses increasingly deploy agents to write code, review documents, and process financial expenses.
The risk
The core danger lies in an agent’s ability to act across connected systems. Microsoft states that an agent’s behavior can shift based on its instructions, the data it processes, or its working context. It might make an error, follow harmful directions embedded in a document, or become compromised. Because these systems operate autonomously, a single unsafe action could expose confidential information or alter business records before any human notices the deviation.
View the original post on X
Who is exposed
Organizations are exposed when they grant agents more permissions than necessary for their specific tasks. For example, an agent designed to review expense claims should only access submissions and company rules, not all the permissions of the employee who created it. If such an agent encounters hidden instructions or attempts an unrelated action, external security controls must limit its scope. The risk escalates when agents inherit broad access levels by default.
Cost and compliance
Microsoft emphasizes that enterprises cannot outsource responsibility for what intelligent systems do on their behalf. CEO Satya Nadella warned that model providers’ safety assurances do not relieve companies of accountability. Security business chief Hayete Gallot added that frontier models cannot be trusted regardless of origin, not because they are malicious, but because they are unpredictable. They are trained to achieve tasks, not to follow organizational norms or judgment.
What should you do?
Managers should audit the permissions granted to each AI agent to ensure they align strictly with the specific task. Implement external monitoring systems that operate independently of the AI model to verify actions before they affect critical data or systems. Regularly review agent logs to detect deviations from expected behavior or unauthorized access attempts.
The background in plain language
Teams usually separate the source of information from the system that acts on it. In this context, the AI model provides suggestions or answers, but a separate security layer decides what data it can touch. Check whether your organization allows agents to inherit broad user permissions. Ensure that controls exist outside the AI lab’s influence to govern actions, limiting damage if the model behaves unexpectedly or follows harmful inputs.
Background: AI agents
AI agents are software systems that perceive their environment, make decisions, and execute actions to achieve specific goals. Unlike standard models, they maintain internal state and interact with external tools, creating unique security risks around tool execution, memory retention, and autonomous decision loops that require specialized controls.
Read the full guide: AI Agents Explained: Answers to Security and Architecture Questions
What to do and how to stay safe
- Review current agent permissions and revoke any access that exceeds the specific task requirements.
- Deploy monitoring tools that operate independently of the AI model to detect anomalous actions.
- Verify that your security infrastructure can distinguish between human and agent activities for accurate auditing.
- Ensure that external security controls, not the AI provider, dictate what actions an agent can execute.
Step-by-step guide: AI Governance Explained: Building Rules for Autonomous Systems
General security guidance from the Vector Tribune newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Why does Microsoft compare AI agents to insider threats?
Microsoft states that agents can access sensitive data and execute actions at high speeds, behaving unpredictably like a rogue human employee if compromised or misinstructed.
How many agent types does Microsoft Entra Agent ID now support?
The company announced support for 35 additional local agent types, bringing the total number of supported types to 60.
What is the purpose of separating intelligence from authority?
This approach ensures that while an AI model provides answers, separate security systems decide what data it can access and which actions it is allowed to take.
