Short answer: The safest way to run AI agents is to manage them like insiders: give each one an identity, grant only the access a task needs, log every action and require human approval for risky ones. That is now mainstream advice. On October 19, 2026, Microsoft launched tools that treat agents as potential insider threats, and Gartner predicted that by 2028, 25% of enterprise security breaches will trace back to AI agent abuse.

Key Takeaways

• Agents are insiders. They work inside your systems with credentials and data access.
• Unpredictability is the risk. Microsoft's Hayete Gallot said frontier models "cannot be trusted" regardless of origin, because they are trained to complete tasks, not to follow organizational norms.
• Separate supply from authority. Satya Nadella argued that the systems that monitor agents should not be controlled by the AI labs that supply them.
• 25% of breaches will trace to AI agent abuse by 2028, according to Gartner.

Why Are AI Agents an Insider Risk?

Traditional security assumes threats come from outside. Agents change that. They log in, read files, call APIs and act on systems with permissions a company granted them. According to TechCrunch, Microsoft's Defender can now detect up to 60 types of agents that employees bring to work on their own, including coding agents. Each unmanaged agent is an account with access and no owner.

Gartner's Gene Alvarez tied the coming breaches to unmanaged machine privileges and legacy credential practices, such as long-lived API keys shared across tools (Gartner).

What Did Microsoft Announce?

• Guardian Agents: Monitors the AI agents a company uses; part of Project Perception
• Agent detection in Microsoft Defender: Finds up to 60 types of agents employees bring to work
• Dynamic access controls: Limits agent access to data and systems and adjusts permissions by task
• Human approval: Required when an agent appears to make a risky decision

Nadella wrote that enterprises need to "separate the supply of intelligence from the authority over it." TechCrunch noted the commercial interest behind the message, since Microsoft sells the monitoring tools, but also that the logic mirrors auditing: oversight should be independent of the party being overseen.

A 6-Layer Security Model for AI Agents

• Inventory: Control: Register every agent with an owner; Question it answers: Which agents are running?
• Identity: Control: Unique identity and short-lived credentials per agent; Question it answers: Who is acting?
• Access: Control: Least-privilege, task-scoped permissions; Question it answers: What can it touch?
• Approval: Control: Human sign-off for high-risk actions; Question it answers: Who approved this?
• Monitoring: Control: Independent oversight of agent behavior; Question it answers: Is it behaving as expected?
• Audit: Control: Logs of every tool call and decision; Question it answers: What happened and why?

How Should Enterprises Roll This Out?

1. Inventory first. List every agent in use, sanctioned or not, with its owner, data access and tools.
2. Give each agent an identity. Retire shared service accounts and static keys for agents.
3. Scope permissions per task. Read-only by default; write access only where the workflow needs it.
4. Put humans on risky actions. Payments, deletions, customer messages and permission changes need approval.
5. Log everything. Record each tool call with inputs, outputs and the human who owns the agent.
6. Test before scaling. Red-team agents with prompt injection and data exfiltration scenarios before production.

Standards such as the Model Context Protocol make it easier to control which tools agents can reach. See our MCP enterprise guide, the Gartner IT Symposium 2026 highlights and why AI agent projects fail.

Frequently Asked Questions

Why are AI agents treated as insider threats?

Agents act inside company systems with real credentials and data access. Microsoft's security chief Hayete Gallot said frontier models cannot be trusted regardless of origin because their behavior is unpredictable, not because they are malicious.

What did Microsoft announce for AI agent security?

On October 19, 2026, Microsoft introduced Guardian Agents to monitor the agents a company uses, agent detection in Microsoft Defender for up to 60 agent types, and access controls that adjust permissions by task and require human approval for risky actions.

How many breaches will involve AI agents?

Gartner predicts that by 2028, 25% of enterprise security breaches will trace back to AI agent abuse.

What is the first step to secure AI agents?

Build an inventory of every agent in use, including tools employees bring in themselves, and give each one its own identity and scoped permissions.

Planning an AI agent or consulting project? Book a consultation with Deployed Labs.